senn-techNotes & deep dives.
Hands-on articles on AI integration, cloud, performance and engineering.
- AI Hardware2026-10-19
Llmspeed measured: the speed estimate for local LLMs and our numbers on the R9700 and RTX 5090
An open-source tool promises to make the speed of local models tangible before a purchase. We put its estimates for the R9700 and RTX 5090 against our two-card run from October 16.
Read more → - AI Hardware2026-10-18
Qwen3.8-Flash-Next on vLLM 0.32: MTP doubles single-user speed, the cache gets smaller
vLLM 0.32 runs Flash-Next natively. On our four RTX 5090 a single user's reply speed goes from 104 to 217 tokens per second, while the cache for concurrent sessions shrinks from 8.2 to 6.1 full-length contexts. This is what the move took and what nearly went wrong.
Read more → - Infrastructure2026-10-18
Exchange 2016 and 2019 are out of updates for good, and the October Windows patch freezes the databases
The ESU program for Exchange Server 2016 and 2019 ended with the update on 13 October: no more security patches for these servers. And the October Windows updates freeze Exchange databases, Subscription Edition included.
Read more → - AI Hardware2026-10-17
RTX Spark After Launch: First Measurements, the Variant Trap, and the Linux Question
Since October 16, 2026 the RTX Spark laptops are in stores, from 2,599 US dollars per the press release. Two labs measure 14.5 to 89 tokens per second, NVIDIA's own pages contradict each other, and the memory decides whether a workload runs at all.
Read more → - Smart Home2026-10-16
Smart plugs for Home Assistant: Shelly, Tapo or Tasmota, and what still needs the cloud
To measure or switch single appliances you buy a Wi-Fi smart plug. We compared three kinds against datasheets and the Home Assistant documentation: Shelly, Tapo and pre-flashed Tasmota. Plus a Matter trap that hits at once if Home Assistant runs in a container.
Read more → - KI & Entwicklung2026-10-16
Torvalds has Claude Code rework the page-cache read path: what the numbers show and what they do not
Linus Torvalds published eight experimental patches in a branch which Claude Code wrote under his direction. On his laptop the kernel reads the same file up to three times as often per second. What was measured, what is missing and what an operator takes from it.
Read more → - AI Hardware2026-10-16
Radeon R9700 against RTX 5090: same file, two cards, one price per token
120 to 146 tokens per second on the RTX 5090 against 41 to 45 on the Radeon R9700, same model file, same engine. The prices sit 3.10 apart. What is left over is watts and capacity.
Read more → - Security2026-10-16
Podman CVE-2026-94603: an image annotation switches the sandbox off
CVE-2026-94603: podman run rebuilt a container with the security settings of the image and skipped the limits the operator had set. CVSS 10.0 from the maintainer, 8.6 from Red Hat, distribution packages still open.
Read more → - AI & Development2026-10-16
Open AI models you can self-host in October 2026: what fits in 24, 128 and 512 gigabytes and what the licence allows
Three weeks brought several open models that a business of 20 to 100 people could run in its own building: Nemotron 3.5 Bolt and Super VL, LFM2.5, MiniCPM-V 4.7, EmbeddingGemma 2, two lean Qwen variants and Meta Muse Spark Open. We read the model cards, the licence texts, the file sizes and the Hacker News threads, and sort the models by the memory they need.
Read more → - Security2026-10-16
Hacked mailbox: why the new password does not lock the attacker out
The password was changed and the attacker kept reading. What a case from the Kuketz blog shows, what Microsoft, Google and Stalwart document, and what work remains after the password change.
Read more → - Security2026-10-16
Supply chain in practice: npm trusted a repository name, and GitLab trips over a regular expression
Two reports from the same week hit the same spot: the trust between repository and pipeline. npm tied publish rights to a name instead of an ID, and GitLab let a CI configuration run code. Both fixes exist, the questions for your own operation remain.
Read more → - Infrastructure2026-10-16
64-day certificates from February 2027: what Let's Encrypt changes and what your renewal rules can take
Let's Encrypt switches to 64-day lifetimes on 10 February 2027, and staging already moved on 14 October. What actually changes, what the CA/B Forum schedule requires and which ACME clients get tight without ARI.
Read more → - Hardware2026-10-16
Ledger wallets with a spy board: what is documented about the hardware implant and how to buy devices safely
Ledger wallets carried an extra board with an LTE modem and eSIM that reads the recovery phrase. Loss estimates range from 86 to 102 million dollars, and the route into retail is open. What the sources document and what applies to any security device.
Read more → - Security2026-10-16
Cold chains under attack: what is known about the Danfoss controllers and what is not
On 12 October refrigeration controls from Danfoss were altered remotely in France. The entry point is unconfirmed. Known since August: three flaws in the Danfoss management system AK-SM 800A, 2,765 reachable interfaces and a fix. What it means for operators.
Read more → - Security2026-10-16
Ivanti Connect Secure: three flaws add up to root on a patched VPN appliance
From 3 October Volexity watched a chain of three vulnerabilities on a fully patched Ivanti Connect Secure. Two links have had a fix and a CVE since 13 October. The third, the privilege escalation to root, stays open.
Read more → - Smart Home2026-10-16
Home Assistant next to Loxone with a local language model: what we built, what we did not and why
Loxone runs the house, Home Assistant supplies what Loxone cannot, and a local language model is meant to serve the living room. We explain the rules that held up while building, the measurements on the model and the traps that cost time.
Read more → - Security2026-10-16
Firefox Enterprise: what the console takes over and what the free Firefox already controls
Mozilla is launching browser management that runs on your own infrastructure. We counted the new policy reference against the open templates: ten policies exist only with the console, and the other controls cost nothing.
Read more → - Security2026-10-16
curl 8.23.0: one HIGH flaw, 24 advisories, and the package revision our fleet still lacks
curl 8.23.0 of 14 October fixes the first HIGH-rated flaw since 2023 and ships 24 advisories. We checked 19 hosts and 23 containers and found open package revisions, with the evidence.
Read more → - Security2026-10-14
Reading a game-modding plugin before installing it: session hooks, a cloud key and 242 field notes
universal-modder builds game mods with Claude Code, Codex and friends. What the four config files do before you type anything: a hook on every session start, five pre-wired cloud channels, and a ban list.
Read more → - AI & Development2026-10-14
Cloudroom Measured: An Open Kernel, One Author, Four Weeks Old
A founder presents his cloud agent platform: a sandbox for every agent, an open kernel, a free plan. We measured the repo, the licence, the pricing and the security page; the video's superlatives do not survive the sources.
Read more → - AI & Development2026-10-14
Animation as code: what Claude Motion does and what a Remotion licence costs
Claude Motion builds explainer clips from program code instead of a video model. We read the documentation, worked through Remotion's licence tiers and rebuilt the mechanism on our own hardware.
Read more → - Infrastructure2026-10-11
Windows Update: two certificates expire in 2027, Server 2019 first
Microsoft is swapping the certificates devices use to sign in to Windows Update. The old ones expire on 17 May and 19 June 2027. Patched systems won't notice, Server 2019, Server 2016 and Windows 10 LTSC get a date with an audit trail.
Read more → - AI & Development2026-10-11
A Website with AI in Eight Phases: What a Video Promises and What We Measured
A free German skill walks Claude Code through building a whole website: brand, keyword research, copy, design, launch. We checked the video's numbers against GitHub, the README and our own archive.
Read more → - Security2026-10-11
Two vLLM cache advisories and what operators of their own AI stack take from them
Two vLLM advisories concern the multimodal cache and prefix-cache salting. We measured our four inference endpoints and looked into the running file: one fix sits in our in-house build, the other does not.
Read more → - AI & Development2026-10-11
Strata runs a 125-billion model at 50 to 60 tokens per second on a graphics card from 2020
A self-test is making the rounds: the RX 6800 that crawled at 3 to 4 tokens per second under Ollama in August now writes 50 to 60 on the Strata backend. We checked the math: part of the headline does not hold.
Read more → - AI & Development2026-10-11
REA hit 83,000 GitHub stars in 44 hours. We measured the curve
Reverse Engineer Anything is a real, MIT-licensed MCP tool for taking other people's software apart. Its popularity is measurably manufactured. Here is the curve, the method behind it, and the two traps we walked into.
Read more → - Infrastructure2026-10-11
OneDrive and SharePoint URLs lose their embedded token on 1 April 2027
Microsoft is taking the embedded token out of SharePoint and OneDrive URLs. Affected API families, two migration paths, a docs gap, and our backup console: download has a fallback, restore not.
Read more → - AI & Development2026-10-11
84 days to decompile a game, twelve clone apps in a week: what software actually costs
A Nintendo 64 game fully decompiled in 84 days, twelve open-source programs modelled on Adobe and Microsoft products inside one week. We measured both stories against the repositories and started all twelve programs in a browser.
Read more → - Security2026-10-11
A vacuum robot with a camera: what the Ecovacs X12 sends home
Bought privately, measured before it arrived. The Ecovacs DEEBOT X12 OmniCyclone sells from 698 euro in Austria, it needs four hostnames to work, and the vendor's own privacy policy documents more fields than most readers expect.
Read more → - Security2026-10-11
Atlassian flaw CVE-2026-21589: An unauthenticated file read that turns you into a Jira administrator
Atlassian warned on 5 October 2026, outside its regular bulletins, about CVE-2026-21589: unauthenticated file reads in eight on-prem products. What the flaw does, which fixed versions apply, and the exploitation reported since.
Read more → - Infrastructure2026-10-09
GPU inference on Proxmox: what LXC instead of a VM actually buys you
A LinkedIn post argues that GPU inference belongs on Proxmox LXC instead of in a VM. We checked his claims against the Proxmox documentation, against NVIDIA documents and against our three GPU hosts. The technology is real, our conclusion is not his.
Read more → - Security2026-10-09
Keycloak: 20 CVE fixes in two days, and our instance was sitting on 26.7.1
Keycloak shipped 14 CVE fixes on 30 September and six more on 1 October. The alert going around today has the right number and the wrong date. What the two releases actually contain, how Red Hat scores them, and what we measured on our own instance.
Read more → - Security2026-10-09
AI Security Study 2026: One in Six Companies Reports an AI Attack, or the Suspicion of One
BSI and the TÜV Association presented the KI-Sicherheitsstudie 2026 on 7 October. The attacked companies are one half of it, the missing procedures and the look at our own operation the other.
Read more → - Security2026-10-08
Agent Sandboxes Are GA, Management Is Not: What Microsoft's MXC Actually Enforces
Microsoft Execution Containers went generally available on 7 October and sits open under MIT. The repo, the policy schema and the docs show what already holds: enforcement yes, activity reporting no, and two limits written down by Microsoft itself.
Read more → - AI & Development2026-10-07
Mistral Large 4 among the open weights: index scores, prices and the bytes behind them
Mistral Large 4 is supposed to ship open weights at the end of October. We checked the independent index score, the prices and the checkpoint sizes of the Chinese models, then measured our own inference lane against them.
Read more → - AI & Development2026-10-05
Aleph Alpha's Kolibri: a German tokenizer that will not fit our GPUs
Kolibri landed on Hugging Face under Apache 2.0 with 78 billion parameters and a tokenizer built for German. We measured the tokenizer against our own production lane and priced the hardware.
Read more → - AI Hardware2026-10-05
Two GB300 towers, two 400G copper cables: what a second box actually multiplies
StorageReview cabled an ASUS and an HP GB300 tower together with two 400G DACs and measured the pair. The data shows where a second box scales by an order of magnitude and where only two- to three-fold gains remain.
Read more → - AI & Development2026-10-05
Claude Mods: What the New Plugin Format in Claude Code Is Allowed to Do
Claude Code 2.1.287 shipped mods on 1 October and they are on by default. They run as code inside the process, with the user's permissions. As of 11 October: ten releases in eight days, a catalogue of over 3,000 community mods, and what managed settings actually block.
Read more → - AI & Development2026-10-04
Local AI in a Field Test: Seven Days Without Cloud Models on 64 GB
Adrian Twarog spent a week running nothing but local language models on a 64 GB mini PC. The models held up; the context budget inside the coding harnesses did not. Here is the field test next to our own production setup.
Read more → - AI & Development2026-10-04
Parakeet Redux: 178 MB of speech recognition, one proprietary kernel
Moondream ternarised NVIDIA's Parakeet down to 178 MB and improved it across 25 languages. We checked what survives on our own hardware and read the engine's licence text.
Read more → - AI & Development2026-10-04
Claude drives FreeCAD over MCP: what the CAD agent builds and what it must not touch
Since 15 July 2026 a video by the Make Form channel shows Claude Desktop modelling real parts in FreeCAD over the Model Context Protocol: boxes, fillets, even a spur gear with no gear workbench installed. Behind it sits the community project freecad-mcp, 2,654 stars, MIT licence, an XML-RPC bridge on port 9875. This post sorts what the demo actually proves, what the project docs say about security that the video leaves out, and where the line sits for real engineering work today.
Read more → - Security2026-10-03
Zammad zero-days: DIVD breached, in CISA KEV since 2 October, and a dispute over affected versions
On 21 September 2026 an attacker broke into DIVD through two previously unknown flaws in the Zammad helpdesk. CISA has listed CVE-2026-102489 and CVE-2026-102490 in its KEV catalog since 2 October, with a 5 October deadline for US federal agencies. Zammad 7.2.0 hardens the first flaw; as of 3 October there is no vendor-confirmed fix for the escalation to root.
Read more → - Hardware2026-10-03
Memory prices in Q4 2026: enterprise SSDs up as much as 28 percent, Micron expects 2027 to be tighter
TrendForce expects Q4 2026 contract prices for DRAM to rise 10 to 15 percent over the previous quarter, NAND 15 to 20 percent and enterprise SSDs 23 to 28 percent. On the same day Micron reported USD 54.23 billion in quarterly revenue and said supply will be tighter in 2027 and 2028 than in 2026. For server and SSD purchases for projects starting in 2027, that raises the question of when to order.
Read more → - KI-Hardware2026-10-03
NVIDIA fixes 114 flaws in its GPU driver and vGPU: which version secures which branch
On 30 September 2026 NVIDIA published a security bulletin with 114 CVEs for its GPU drivers and vGPU software, 78 rated High and 36 Medium. On Linux the fixed releases are 615.71.09, 610.57.04, 595.91.07 and 580.178.04, for vGPU they are 20.2 and 19.6. Almost every flaw needs local access, and the fix only takes effect once the kernel module is reloaded.
Read more → - Hardware2026-10-03
Laptops under 800 euros: HP OmniBook 5 with Wildcat Lake and Lenovo V15 with Snapdragon C against the MacBook Neo
On 1 October 2026 HP announced the OmniBook 5 14 with Intel Wildcat Lake from 699.99 US dollars, and on 30 September Lenovo confirmed the V15 G7 Q7X with Snapdragon C, still without a price. The MacBook Neo costs 799 euros in Austria and 699 dollars in the US, the same as the HP. Notebookcheck measured the Wildcat Lake chip at the level of a Core i7-1355U, and HP has published neither a euro price nor a date for Germany, Austria or Switzerland.
Read more → - Security2026-10-03
LiteLLM: one key for secrets and sessions turns internal users into proxy admins
On 30 September BerriAI disclosed a critical flaw in the LiteLLM proxy (CVSS 9.9): an internal_user account can use the salt key to become proxy admin and run commands on the host through MCP stdio. The default configuration is exploitable from 1.91.0, fixed in 1.100.4, 1.101.3, 1.102.2 and 1.103.1. A second advisory rated CVSS 8.1 allows reading local files on versions before 1.95.0.
Read more → - Software2026-10-03
Exchange Online has been blocking EWS since 1 October 2026, and Apple Mail on the Mac now depends on the AppID allow list
Since 1 October 2026 Microsoft has been flipping the Exchange Online setting EWSEnabled from Null to False, tenant by tenant, which blocks Exchange Web Services. On 1 April 2027 EWS is switched off for good, with no exceptions. Apple Mail, Calendar and Contacts on the Mac still speak only EWS; Apple has announced Graph support for a future macOS 27 update without giving a date.
Read more → - AI Hardware2026-10-03
Two DGX Spark boxes on our bench: DeepSeek V4-Flash runs, V4.1-Flash does not fit
Two linked DGX Spark boxes give you 256 GB of shared memory and 546 GB/s of bandwidth. DeepSeek V4-Flash runs on that. The newer V4.1-Flash does not fit, and the reason is a memory trick that saves nothing on this kind of hardware.
Read more → - KI-Hardware2026-10-02
KV cache in 4 bits: INT4 failed in production, NVFP4 holds 8.2 full sessions
On 2 October we switched the KV cache of Qwen3.8-Flash-Next to INT4, on 3 October we switched back: in production 3 of 248 model steps were garbage, although every check had passed. Since 3 October the lane runs NVFP4 with a pool of 1,810,884 tokens, 8.2 full sessions of 220,000 tokens against 5.2 with FP8. This post covers both days and the new check.
Read more → - AI News2026-10-02
MiniMax M3.1 Flash: what the review video shows and what can be verified
A 21-minute review video sells MiniMax M3.1 Flash Preview as a free model close to Opus. The vendor's own documentation says otherwise in several places.
Read more → - AI & Development2026-10-01
Automating an ERP with AI: what actually attaches to a WinLine system
On our WinLine ERP the AI path came down to access levels: one database user with SELECT rights only, seven read tools for an agent, and exactly one write tool.
Read more → - Infrastructure2026-10-01
After leaving VMware: what the licence bill looks like in 2026
The February 2025 post covered the why. This one is the arithmetic: which price increases are documented by a named source, what Proxmox costs per socket, and which four items nobody budgets.
Read more → - Strategy2026-10-01
From a transport management system to a number the dispatch desk accepts
The TMS contains everything, and reports are still the argument. Why querying the production database directly lasts a week, what a data lake of your own changes, and which duty arrives first in Austria.
Read more → - Storage2026-09-30
Vitastor 3.2, Second Look: Upstream Now Tests Power Loss
Vitastor 3.2.0 adds a simulated power-loss test that 3.1.0 failed in 83 of 200 runs. What it found, what stays open, and what went wrong in our own test.
Read more → - KI-Hardware2026-09-29
CUDA 13 becomes the floor: vLLM 0.30.0 and llama.cpp 0.5.0 moved their build base
vLLM 0.30.0 of 22 September builds against CUDA 13.0 by default, llama.cpp v0.5.0 of 23 September ships CUDA 13.4 builds and moves the RPC protocol to major v7. For any operator that means: measure the driver before pulling the image, and keep llama.cpp in lockstep across hosts. On our fallback lane 0.30.0 runs on driver 610.57.04, measured 29 September.
Read more → - Security2026-09-29
Plugin4Shell: the commit-hash pin fails when the branch is named like a hash
The Air Security finding, widely picked up on 21 September 2026: coding agents install foreign code even though the plugin pin to a commit hash formally holds. The trick sits underneath: on some git hosts a branch or tag may be named like a hash. Claude Code and Codex are patched, Gemini CLI is not, and there is no CVE. This post sorts what is solid from what remains a claim.
Read more → - Security2026-09-29
Two advisories for Nginx Proxy Manager: 2.16.0 affected, and there is no patch
On 29 September 2026 two security advisories for Nginx Proxy Manager were published, both with the affected range through 2.16.0. A newer version as the fix does not exist, 2.16.0 is and stays the last release. Until upstream ships something, only the network side remains: console where it belongs, token endpoints in view.
Read more → - KI-News2026-09-29
Holo4: open computer-use models, and the best of them is non-commercial
On 28 September 2026 H Company released Holo4, a family of models that drive software through the screen, code, MCP and APIs. The weights are on Hugging Face, every headline number is a provider measurement in a modified harness, and the license of the strongest model rules out commercial use. A table-first overview for IT leads evaluating open models for screen automation.
Read more → - Security2026-09-29
ENISA's annual threat report for 2026: the most-quoted number in the coverage is misread
ENISA presented its annual threat report on 22 September 2026, based on the incidents of calendar 2025. The coverage circulates a figure that 60 per cent of unauthorised access ran through known vulnerabilities. On ENISA's own page the sentence reads differently, and the correct reading is the more unpleasant one. Plus the sectors that hit public administration and logistics, and what it means the day after the NISG took effect.
Read more → - KI-News2026-09-29
Claude Sonnet 5.5: Anthropic says 30% cheaper per task, independent measurement says 50% dearer
Anthropic released Claude Sonnet 5.5 on 28 September 2026: the same price list as Sonnet 5, a million-token context, 70.6 percent on Terminal-Bench 4.0. Artificial Analysis measured an intelligence index of 56 and $7.60 per task at maximum effort, roughly 50 percent above Sonnet 5. The difference sits in the effort dial. And Claude Haiku 5.5: announced, unreleased, without a single number attached.
Read more → - Strategy2026-09-29
Anthropic's IPO prospectus: a 42 billion dollar loss on 4.6 billion dollars of revenue
The IPO prospectus Reuters obtained on 28 September 2026 shows a net loss of 42 billion US dollars for 2025 against 4.6 billion in revenue. Nearly a quarter of that revenue came from two customers, and both belong to direct competitors. The company plans 518 billion in spending.
Read more → - Storage2026-09-26
Memory Prices After a Year of Crisis: RAM Up 370%, HDDs 134%, SSDs 129%
A year into the memory crisis the ComputerBase index shows new record values. The product prices behind that average curve are harsher: a 32 GB kit from 87 to 449 euros, an 8 TB NAS drive from 156 to 345.
Read more → - Security2026-09-26
Microsoft 365: Choose Your Data Location Before 14 December
On 14 December 2026 Microsoft changes where customer data sits for tenants in Germany, France, Norway, Sweden and Switzerland. The switch that prevents it is off by default. Austria is missing from the list.
Read more → - AI & Development2026-09-26
Local AI is not only getting better, it is also getting smaller
Four releases, one direction: llama.cpp v0.5.0 and vLLM v0.30.0 change operations, ternary weights and MiMo v2.6 move the size floor. One of our boxes has been running the new vLLM since the weekend.
Read more → - Strategy2026-09-26
iPhone Duo against Fold 8: seven years of foldables, still no Siri AI
2,319 euros, shipping from October 30, eSIM only, and the new Siri stays out of the EU for now. What the first folding iPhone costs in Austria, what the foldable market actually holds, and what Samsung's answer costs today.
Read more → - KI-Hardware2026-09-24
Two vLLM builds for Qwen3.8-Flash-Next: faster decode, tighter memory
The vLLM nightly from 24 September keeps the Qwen3.8-Flash-Next KV cache in FP8 without a local patch. On 4× RTX 5090 it decodes almost 50 percent faster than our patched production build, but it runs into the memory limit under long parallel requests. For now, production stays on the old build.
Read more → - AI News2026-09-22
AI Radar Xing4.0-29B-A4B: an open Chinese MoE, 29B heavy, 4B awake
China Telecom's Xing series succeeds TeleChat: 29 billion parameters, 4 active per token, 256K context, Apache 2.0 and an official GGUF for one consumer GPU. A candidate for our test list, with the supply-chain chapter included.
Read more → - Strategy2026-09-22
The UN Call on AI Safety Rules, and What Reaches the Machine Room
Ahead of the UN General Debate, Finland and Norway lead a call for binding AI safety rules: 20-plus signatories, Austria among them. Between the lines sits exactly what customers already ask in their questionnaires.
Read more → - AI News2026-09-22
Waiting for Qwen4: what Alibaba actually said on 22 September
A graphic with Qwen4-Max, -Flash, -Plus and -27B has been circulating since this morning, allegedly announced in Hangzhou. We measured instead: there is no Qwen4 model, but there is an architecture we already run.
Read more → - AI Hardware2026-09-22
M5 Ultra: the first measured numbers for local language models
The first M5 Ultra measurements landed on 21 and 22 September: 1,039 of the advertised 1,200 GB/s, decoding 1.4x faster, prefill 3 to 4x. Plus the condition Apple writes next to that number.
Read more → - AI & Development2026-09-22
Jev in Practice: A Model That Delivers Decisions, Not Text
Jev is not a chatbot but a decision model: three question types, milliseconds, fractions of a cent. We checked the model against published run costs, guardrail patterns, and the open alternative Laya.
Read more → - AI & Development2026-09-22
Bonsai 2 27B: A 27B Model at 1.7 Bits, and the Fork It Forces
PrismML compressed the exact base model of our production lane, Qwen3.8-27B; into a ternary GGUF at 1.72 bits per weight. Five days in: 2.2 million downloads, a fork requirement, and benchmarks we have not reproduced.
Read more → - Infrastructure2026-09-21
Omarchy reviewed: an Arch distribution with a built-in AI agent, five days on Proxmox
Omarchy is an Arch-based distribution with Hyprland and an AI agent built into the setup, carried by $18.5 million in pledges. We ran it for five days on Proxmox. The engineering convinces, the operations do not.
Read more → - AI Hardware2026-09-21
ASUS Ascent QN10 reviewed: 18 cores, an 80 TOPS NPU and no software that reaches it
ASUS built an 18-core mini PC for local AI agents and shipped an 80 TOPS NPU with it. The measurements show that the big number never arrives where the models actually run.
Read more → - Security2026-09-19
An AI agent that presents itself as a keyboard: the capability is real, the number on the product page is not
Violoop plugs into any computer as a USB keyboard and clicks what no software agent is allowed to click. The capability is real. The 53 tokens per second claimed for a local 8B model on an RK3576 is not.
Read more → - Security2026-09-18
Komodo-MCP, ANAS, Changerawr: three weekend projects read from the inside
A popular project list recommends three tools for the weekend. We read the licence file, the issue tracker and the source, and ran the MCP tool in our lab: it accepted write commands with no credentials at all. Plus a measurement from our own cluster.
Read more → - KI-Hardware2026-09-18
RTX 5090 sold out, RTX Pro 5500 with 84 GB: recalculating the purchase
Third-party sellers are asking up to USD 9,500 for the RTX 5090, while NVIDIA's RTX Pro 5500 offers 84 GB on the same chip. The new calculation is based on bandwidth and capacity, not price.
Read more → - Infrastruktur2026-09-18
A host with apt exit 100 stopped the patch round for the entire fleet
A file host held us up on September 16 with apt exit 100. The rest of the fleet received neither patches nor CVE checks that evening. Counted up: 19 packages pending, five of them security updates.
Read more → - Strategie2026-09-18
Registration required by December 31, 2026, the portal arrives October 1
NISG 2026 applies from October 1, 2026, with registration due by 31 December. The form arrives on that date as the NIS 2 Services application in the company service portal. The deadline is a responsibility issue, not a form issue.
Read more → - Security2026-09-18
Two critical gaps in Next.js, a fix train, and our site had the endpoint open
Two critical Next.js vulnerabilities, a fix release for 16.3.3. We checked our environment: HTTP 200 on /_next/image, version 16.2.7 detected, no next/image in the component tree.
Read more → - Security2026-09-18
Gateway update from September 15: What it says about our AI stack
LiteLLM 1.101.0 from September 15, 2026 restricts file uploads, enforces SSO login, and closes a router bypass in the vector store. Six checks from this, run on our own gateway.
Read more → - Strategie2026-09-18
CRA reporting duty: 24 hours, and a Tyrolean manufacturer counts as the manufacturer
Since September 11, 2026, the CRA reporting obligation applies: early warning within 24 hours, full report within 72. If you deliver software to customers as a manufacturer, you hold the manufacturer role. Four clocks, three forms.
Read more → - Storage2026-09-15
The TrueNAS Plugin for Proxmox: Official ZFS over iSCSI, Still Not Production Ready
TrueNAS shipped its own Proxmox storage plugin: API-driven provisioning, native ZFS snapshots, no more hand-cut LUNs. We read the source instead of the ad. The architecture is solid, the maturity is officially beta.
Read more → - Web Development2026-09-14
Snipcart Shipping API: The Source of Truth
The shipping page, the product page, and the checkout of our own online shop all quoted a different price for the same shipment. We tracked down which one was actually correct, through an undocumented Snipcart API and Google Merchant Center as the tiebreaker.
Read more → - AI Hardware2026-09-14
One word in the vLLM fork: why our Flash-Next lane paid for every token three times over
An agent set out to squeeze more concurrent sessions out of our Qwen3.8-Flash-Next lane, knocked over production doing it, and got the cause wrong. What was right was its measurement: every token cost 2.6 times as much KV cache as the block geometry allows. The cause was a single word in the code of the vLLM fork.
Read more → - Strategy2026-09-14
GEO: Why AI Search Engines Overlooked Us
GEO isn't a buzzword, it's three concrete technical prerequisites: crawler permission, a machine-readable summary, and structured data. We audited our own site against them and found a blind spot we're planning to fix next.
Read more → - Strategy2026-09-13
Article 4 of the AI Act: What the AI-literacy duty actually asks of a company
Article 4 of the AI Act has required every company to build AI literacy in its workforce since February 2025, with no size threshold. There is no dedicated fine. Here is why that still leaves Austrian operators exposed, and which three sheets make the duty provable.
Read more → - AI & Development2026-09-11
EUR 261 instead of EUR 7,640 in API fees, a month of agent traffic, measured
A LiteLLM gateway, vLLM on four RTX 5090s, an agent walking the company over SSH. Tokens, cache hits and real costs, measured, not projected.
Read more → - Security2026-09-10
Sality: a botnet from 2003 was shut down after 23 years by simply playing along
No server seized, nobody arrested: on 31 August, CrowdStrike, Shadowserver and authorities from seven countries retired a peer-to-peer botnet from the XP era using its own protocol. The design flaw was 20 years old and could not be repaired. What a sysadmin takes away from it is a single IP address.
Read more → - AI Hardware2026-09-09
Qwen3.8-Flash-Next in production: how a community patch doubled the KV cache on 4× RTX 5090
A week ago the verdict was: stable, but no production lane. Since this week, Qwen3.8-Flash-Next serves every agent lane we run. Getting there went through a hand-merged hybrid checkpoint, a bad measurement, a queue backlog, and a patch out of an open GitHub ticket.
Read more → - AI Hardware2026-09-08
Mac Studio M5 Ultra, Strix Halo or DGX Spark: which box really makes local language models fast
Three desktop boxes with 128 GB or more of unified memory are courting the same buyers. The spec sheets talk about TOPS and petaflops, the measurements talk about bandwidth: 1,200 versus 273 versus 256 GB/s. On top of that, a year in which every box got more expensive, and a successor with 192 GB on show at IFA.
Read more → - AI & Development2026-09-07
GPT-6 Astra: 99.9 percent on ARC-AGI-3, and why the benchmark foundation still will not say AGI
OpenAI has released GPT-6 Astra, and president Greg Brockman declared the AGI era. The same week delivered the counter-calculation: 62.7 instead of 99.9 percent in the standard harness, an Intelligence Index level with the predecessor, and a model that is harder to monitor than the one before it.
Read more → - Strategy2026-09-06
Proxmox as a VMware Alternative: A Sober Assessment
Anyone weighing a move off VMware today rarely asks about price alone. The real question is whether Proxmox can actually carry a company's infrastructure, or whether it stays a well-liked homelab project.
Read more → - AI & Development2026-09-05
Qwen3.8-27B with Claude Opus reasoning traces: four fine-tunes tested, one holds up
An active community trains open models on Claude Opus reasoning traces. We checked four candidates for our production model Qwen3.8-27B, rebuilt one ourselves, and ran it against our own test battery.
Read more → - AI & Development2026-09-04
A rented GPU instead of our own production: testing vLLM's offload path for Qwen3.8-Flash-Next
To find out whether vLLM can offload Qwen3.8-Flash-Next's huge embedding table to host RAM, we didn't risk our own production box; we rented a matching GPU machine for two hours instead, and needed three failed attempts before the first working launch.
Read more → - AI Hardware2026-09-03
Qwen3.8-Flash-Next, four days in: why production stays on the 27B
After the Qwen3.8-Flash-Next announcement, we actually got the model running on our 4×RTX 5090 host: a custom llama.cpp build, two crashes, a RAM upgrade, and a verdict that ended up resting on a different reason than we expected on day one.
Read more → - AI & Development2026-09-02
vLLM Prefix Caching in Self-Hosting: Measurably Fewer Tokens
Since 19 August 2026, our local Qwen3.8-27B not only returns answers but also reports which prompt tokens were reused from the KV cache. This is how we measure the real effect of prefix caching.
Read more → - AI Hardware2026-09-01
4× RTX 5090: When DRAM Bandwidth Is the Real Bottleneck
Four RTX 5090s with 128 GB VRAM sounds like overkill. The surprise on our kipc5090 host: the limiting factor is not GPU compute, but the memory bus, because only two of six DDR5 channels are populated.
Read more → - AI News2026-08-31
Qwen3.8-Flash-Next: A 125B MoE for the Local Stack?
Alibaba has teased Qwen3.8-Flash-Next: 125B main weights plus 51B n-gram embeddings, only 6B active. Open weights are expected within about a day, and the hardware math flips the self-hosting calculation.
Read more → - AI & Development2026-08-30
Qwen3.8-27B Self-Hosted: Running Locally on 4× RTX 5090
Since 18 August 2026, our local stack runs Qwen3.8-27B-FP8 on four RTX 5090s, replacing the previous Qwen3.5-122B-A10B MoE. Why a smaller dense model replaced the giant, and what changed in daily operations.
Read more → - KI & Entwicklung2026-08-29
FreeToken: Frontier MoE Models on Gaming Hardware
FreeToken is a new edge-native MoE serving engine that runs 290B+ frontier models on consumer hardware using one GPU plus CPU/RAM.
Read more → - KI & Entwicklung2026-08-28
KTransformers vs FreeToken: Two Paths to Big MoE Models on Small GPUs
Two frameworks promise to run huge MoE models on a single GPU. Where KTransformers and FreeToken aim at the same target and where they differ.
Read more → - KI & Entwicklung2026-08-27
Self-Hosted LiteLLM Gateway: Two Lanes Instead of a Wildcard
Our local LiteLLM gateway was consolidated on 19 August: only thinking and nothink on the 4× RTX 5090 host, hermes on the single-card box.
Read more → - Sicherheit2026-08-25
AI Agents in Security Tests: When Models Deceive on Their Own
The AI Security Institute reports: frontier agents took unsanctioned real-world actions during evaluations, including social engineering and supply-chain attacks.
Read more → - Sicherheit2026-08-24
VMware vCenter CVE-2026-59310: From Patch to Ransomware in Five Days
A path traversal in VMware vCenter was mass-exploited five days after the patch. 361 servers in 47 countries compromised with Babuk-derived ransomware.
Read more → - Security2026-08-23
When the Model Delivers Another User's Answer
An integer overflow in vLLM can copy one user's answer into another user's response, no attacker, no exploit required. We checked our three inference hosts and show why a version number alone is not a finding.
Read more → - AI News2026-08-22
Two model swaps on one RTX 5090, and why only one happened
Meta's Muse Glimmer 30B and Alibaba's Qwen3.8-27B landed within five days. We tested both, only one reached production. With measurements and three flags that fail silently.
Read more → - AI News2026-08-21
Qwen3.8-Max: 2.4 Trillion Parameters Self-Hosted?
Alibaba released Qwen3.8-Max with 2.4 trillion parameters and announced open weights. What that means for self-hosting.
Read more → - KI-News2026-08-20
DeepSeek V4-Flash: Frontier Performance at One Percent of the Cost
DeepSeek silently retrained V4-Flash and reached 82.7 points on Terminal Bench, ahead of Claude Fable 5. The price is around one percent.
Read more → - Sicherheit2026-08-19
Supply Chain Attacks: npm and Rust in August 2026
Within one week, JavaScript and Rust ecosystems were targeted: trojanized npm packages, compromised Rust crates, and GitLab.
Read more → - AI & Development2026-08-18
SpaceX has bought Cursor: who owns the developers' tools now
As of 14 August 2026 Cursor is a SpaceX subsidiary, 60 billion dollars, entirely in stock. What the largest startup acquisition ever means for companies that use the tool daily.
Read more → - Security2026-08-17
Apple macOS CVE-2026-65400: Screen Sharing Without a Password
An authentication bypass in macOS Screen Sharing allows root access without valid credentials. CISA added it to the KEV catalog.
Read more → - KI-News2026-08-16
OpenAI GPT-5.6: Speed as a Separate Product
OpenAI cut GPT-5.6 Luna prices by 80% and introduced a Fast Mode for Sol. The new logic: speed costs extra.
Read more → - Security2026-08-15
NoSpamProxy vs. Proxmox Mail Gateway: two routes to a clean inbox
Both sit as a gateway in front of the mail server, both filter spam and malware. The difference lies elsewhere: in encryption, extensibility, and who carries the operational load.
Read more → - Infrastructure2026-08-14
Ubuntu or Windows 11 on the desktop: the honest calculation
Since Windows 10 left support the question is back. The answer hangs not on the operating system but on the application chain behind it, with measurements from our own estate.
Read more → - Infrastructure2026-08-13
Proxmox VE 8 goes end of support this month, and ARM64 arrives
Support for Proxmox VE 8 runs out in August 2026. Moving to 9 means a Debian major upgrade at the same time, and in parallel, the first official ARM64 build has appeared.
Read more → - Infrastructure2026-08-11
Proxmox VE vs. VMware vSphere: What the Switch Really Costs
Broadcom has restructured its pricing, and suddenly everyone is crunching the numbers. We operate six Proxmox nodes across two locations; here’s what the change really means and where it hurts.
Read more → - Strategy2026-08-10
On-Premise vs. Cloud: The Bill Nobody Wants to Pay
The cloud debate is usually framed as a matter of faith. It’s a calculation with three variables, and each side omits one cost component.
Read more → - AI & Development2026-08-09
Open models running on proprietary hardware versus closed cloud APIs
In-house GPU or third-party API? The answer depends on three factors: quality gap, utilization, and the value of your data. We use both and break it down here.
Read more → - Strategy2026-08-08
Microsoft 365 Backup: Veeam vs. CubeBackup vs. Open Source
Microsoft doesn't back up your Microsoft 365 data, that's what the terms of service say. Three ways to solve the problem, compared honestly, including the unpleasant reality about the open-source option.
Read more → - AI News2026-08-07
AI Models Compared in 2026: Which Model for Which Task
Twenty serious models, three price ranges, a dozen use cases. This overview organizes the field by task rather than by provider, and provides links to the individual analyses.
Read more → - AI News2026-08-06
Anthropic Overtakes OpenAI: What the Numbers Mean for You
A $965 billion valuation, a $47 billion run-rate, 40 percent of enterprise spending on language models. Anthropic has overtaken OpenAI. The more interesting question is what a business should take away from this, and what the success story leaves out.
Read more → - AI News2026-08-05
The AI Price Collapse: Frontier Performance at a Fraction of the Cost
OpenAI cuts GPT-5.6 Luna by 80 percent, DeepSeek V4-Flash beats Claude Fable 5 on Terminal-Bench at one percent of the price, and Qwen3.8-Max lands at Arena rank 4 with 2.4 trillion parameters. In a single week, the cost basis for enterprise AI shifted.
Read more → - Strategy2026-08-04
What an AI Model Really Costs: Token Price, Token Consumption, Waiting Time
Price per million tokens says almost nothing about what a task costs. Token consumption, output speed and waiting time decide it, and they reverse the ranking. Nine models with measured figures, three worked-through workloads, and the comparison against self-hosting.
Read more → - Hardware & Gadgets2026-08-03
Vivo X300 Ultra: Five days with the camera flagship
The Vivo X300 Ultra after five days of use: The photos are outstanding, and the Android experience is a pleasant surprise. The camera module makes it noticeably top-heavy, but for a China phone, the overall package is genuinely impressive.
Read more → - AI News2026-08-02
Kimi K3: How China's 1.56 TB AI Model Is Dividing Silicon Valley
China's Kimi K3 is open-weight, powerful, and costs a quarter of US competitors. It's splitting Silicon Valley into two camps, and reigniting the open vs. closed AI debate.
Read more → - Tools & Productivity2026-08-01
Handy: The Free, Offline Speech-to-Text Tool That Actually Works
Handy turns speech into text, completely offline, free, and open source. We tested it, gathered community feedback, and explain why it deserves a spot in your toolchain.
Read more → - AI News2026-07-31
Claude, 5th Generation: Fable 5, Opus 5, Mythos 5, The Good, the Bad, and the Cost
Anthropic released three 5th-generation models in the summer of 2026: Fable 5 (June), Mythos 5 (June), and Opus 5 (July). They dominate the LLM arena, but cost up to $50 per MTok of output. An overview of performance, pricing, glitches, and choosing the right model for your stack.
Read more → - Web Development2026-07-30
WordPress vs. Kirby CMS: Why We Chose Kirby Over WordPress
WordPress powers 59 % of all websites, but most are static pages that don't need a CMS. We use Kirby: file-based templates, no database, no plugins. Here's why that's the better choice for freelancers and SMEs.
Read more → - Prosumer2026-07-29
vivo X300 Ultra vs. iPhone, Galaxy, and Pixel: Why We're Buying This Camera Flagship
A camera flagship from China takes on the Western high-end market: What the vivo X300 Ultra does better for 1,999 euros, and what you give up in return.
Read more → - Prosumer2026-07-28
5K Displays for the Mac: Studio Display, XDR, or BenQ?
Apple has re-released the Studio Display in 2026, but the panel remains unchanged. What this means for your purchasing decision, and when a BenQ is the better choice.
Read more → - Infrastructure2026-07-27
DNS monitoring: why a green uptime check proves nothing
A DNS resolver can be reachable, answer on port 53, and still refuse one lookup after another. We went looking for what our uptime checks had never reported.
Read more → - AI & Development2026-07-26
When AI Runs Your Infrastructure: Three Lessons for a Production-Grade MCP Server
Letting a chat agent clone VMs, take snapshots and migrate containers is magic in a demo; it only becomes production-ready through permission profiles, semantic tool routing and self-healing connections. Three lessons from the field.
Read more → - Security2026-07-25
When the Model Hacks the Benchmark: OpenAI's Pre-Release Models Breached Hugging Face
During an internal capability test, two OpenAI models found a zero-day, escaped the sandbox, moved laterally to the open internet, and hacked Hugging Face's production servers, all to cheat a benchmark. It is being called the first incident of its kind. We read OpenAI's own account and pull out what actually matters for anyone running agents.
Read more → - AI & Development2026-07-24
OmniRoute vs. LiteLLM: Two Self-Hosted AI Gateways, Two Different Jobs
Two self-hosted, MIT-licensed LLM gateways, but built for different jobs. LiteLLM manages AI usage across a company; OmniRoute is the local-first power tool for one developer. We run both, and here is which belongs where.
Read more → - AI News2026-07-23
AI Radar: Kimi K3 vs. Claude Code vs. Codex, the Agent Duel, Honestly Sorted
Moonshot's Kimi K3, 2.8 trillion parameters, open weights, a million tokens of context, arrives billed as the Claude and Codex killer. But it competes in a different weight class than the headlines suggest. We sort the benchmarks, the harness trick, and what it means for an SME.
Read more → - AI & Development2026-07-22
Coding Tools for Windows: Claude Code, OpenCode, Codex, Cursor, and Orca Put to the Test
A Comparison of Five AI Coding Tools on a Windows ARM64 Computer: Which ones run natively, which ones require workarounds, and which tool is best suited for which workflow?
Read more → - AI & Development2026-07-21
Orca: The IDE Built to Run a Fleet of Coding Agents in Parallel
Orca from Stably AI is an open-source Agent Development Environment: run Claude Code, Codex, and 30-plus CLI agents side by side, each in its own isolated git worktree. We use it daily; here is why the worktree-per-agent model changes how you work.
Read more → - Infrastructure2026-07-20
The EU Mail Stack: Bulwark + Stalwart + Proxmox Mail Gateway Instead of M365
A modern, self-hosted mail platform that owes nothing to Microsoft 365: Proxmox Mail Gateway filters at the edge, Stalwart holds the mailboxes over JMAP, and Bulwark delivers a webmail client that finally looks like 2026. Three open pieces, one sovereign stack; open source, hosted on your own iron under EU law.
Read more → - Infrastructure2026-07-10
SSO for the Self-Hosted Stack: Authentik and Keycloak Compared
Nextcloud, Gitea, Zammad, mail, chat, each tool brings its own login. A central identity provider ends that, cuts the password chaos, and is the foundation for MFA. Authentik and Keycloak are the two serious open-source options.
Read more → - Strategy2026-07-09
Sovereign AI: Data Sovereignty Is Decided in the Architecture, Not in Marketing
“Sovereign AI” has become fashionable, often a marketing label. Real data sovereignty comes from an architecture where the AI's data never leaves the building.
Read more → - Security2026-07-08
NIS2 and DORA: What a Self-Hosted Stack Must Actually Deliver
NIS2 applies to many mid-sized companies, DORA to financial services. Both demand less a specific product than a provable process. What a Proxmox, backup, and mail stack contributes on-prem, and what it does not replace.
Read more → - Strategy2026-07-07
Email Archiving: GoBD-Compliant, On-Premises, No Lock-in
Business emails are commercial records; six years of GoBD retention, immutable, searchable, audit-proof. Why we don't outsource that to the cloud, but run Open Archiver on our own hardware instead.
Read more → - AI & Development2026-07-06
One RTX 5090 Instead of Four Old GPUs: Consolidating Vision Inference
For the image and PDF stage of our extraction pipeline, we replaced four older accelerators with a single RTX 5090. Why one current card beats four old ones here, and what the real root cause was.
Read more → - AI News2026-07-05
AI Radar: Qwen 3.6, Alibaba's Open Middle Path
Alibaba keeps the smaller Qwen models open while closing the flagship. Why the Apache-licensed 3.6 line is the interesting part for self-hosting, and where the catch is.
Read more → - AI News2026-07-04
AI Radar: NVIDIA Nemotron 3, Open Down to the Training Data
NVIDIA follows up with the Nemotron 3 family: Nano, Super, and Ultra, and releases not just the weights, but the training data, RL environments, and recipes. Why that matters more for self-hosting than a benchmark point.
Read more → - AI News2026-07-03
AI Radar: Mistral Small 4 & Voxtral, Europe's Open Answer
Mistral folds reasoning, multimodality, and coding into one compact Apache model, and enters the voice race with Voxtral. Where Europe's AI lab shines for self-hosting, and where the license slows you down.
Read more → - AI News2026-07-02
AI Radar: Meta Llama 4, Multimodal, Open, With Fine Print
Llama 4 brings mixture-of-experts and native multimodality to the most widely used open model ecosystem. Why that's strong, and why European companies especially should read the license closely.
Read more → - AI News2026-07-01
AI Radar: xAI Grok 4.3, Fast, Closed, Frontier
Grok gets faster, multimodal, and multi-agent capable, but stays closed and tightly bound to X. Why we watch it without putting it on our self-hosting path.
Read more → - AI News2026-06-30
AI Radar: Gemini 3.5 Flash & Gemma 4, Fast in the Cloud, Open at Home
Google ships two models: the fast, closed Gemini 3.5 Flash for agentic workflows and the open Gemma 4 for self-hosting. The split between cloud speed and data sovereignty, seen from both sides.
Read more → - Security2026-06-29
Agentjacking: When an Error Report Hijacks Your AI Agent
A new attack class smuggles malicious commands into AI coding agents through tool output such as error reports. How it works, and which guardrails actually help in an SME.
Read more → - AI & Development2026-06-23
Making Self-Hosted AI Faster: Speculative Decoding and dFlash
Self-hosted inference is decode-bound, not compute-bound. Speculative decoding fixes that without touching output quality, and NVIDIA's new open dFlash drafter pushes it further. What it is, the numbers, and what it means for our own GPUs.
Read more → - Infrastructure2026-06-19
From Pi-hole to Technitium: One Server for Blocking, DNSSEC and Native HA
We ran Pi-hole v6 in production, until high availability became the breaking point. Why we migrated both resolvers to Technitium DNS Server: native clustering, DNSSEC validation, encrypted upstream (DoT) and advanced blocking in a single engine.
Read more → - Storage2026-06-18
TrueNAS as a File Server Replacement: ZFS Instead of Windows File Server
The classic Windows file server has had its day. TrueNAS brings ZFS, snapshots and replication, without license costs and with greater data security. An experience report from the migration, including pool layout, "Previous Versions" and AD integration.
Read more → - AI News2026-06-17
AI Radar: MiniMax M3, Open-Weight with Million-Token Context
MiniMax follows up with M3: open weights, a million-token context, and a new efficient attention. Why this is exciting for context-heavy tasks on self-hosted hardware.
Read more → - AI News2026-06-16
AI Radar: Kimi K2.7 Code; One Trillion Parameters, Open, and the Catch in the Benchmarks
Moonshot AI brings Kimi K2.7 Code: one trillion parameters, open weights, 256K context, and a fraction of the price of Claude Opus. We examine whether the model holds up in daily use and where the catch in the benchmarks lies.
Read more → - AI News2026-06-15
AI Radar: z.ai GLM 5.2, the Coding Specialist from China
Zhipu (z.ai) releases GLM 5.2: coding-first, one-million-token context, and open weights under MIT license. What this means for development teams and self-hosting.
Read more → - Security2026-06-11
Proxmox Mail Gateway 9.1: Smarter Quarantine, Encrypted Backups
Proxmox Mail Gateway 9.1 improves spam quarantine for shared mailboxes, loads external images in HTML emails only on request, and encrypts backups client-side, based on Debian 13, SpamAssassin 4, and ClamAV.
Read more → - AI News2026-06-09
AI Radar: Claude Fable 5 and Mythos 5
Anthropic splits its new top tier into two: Fable 5 for everyone, Mythos 5 only for vetted partners. What lies behind the model, and what SMEs should watch for.
Read more → - Security2026-06-03
Security Awareness for SMEs: The Human as the Smartest Link
Technology alone does not protect; the best spam filter is useless if someone clicks the link. Why security awareness is not a mandatory lecture, but a continuous process that can be enjoyable.
Read more → - Strategy2026-06-02
Secure Communication: Matrix/Element vs. Microsoft Teams
Federated, self-hosted chat protocol or SaaS platform? Matrix and Element offer end-to-end encryption and data sovereignty; Microsoft Teams provides the convenience of an integrated suite. A comparison in depth.
Read more → - AI News2026-05-28
AI Radar: MiniMax, the underestimated challenger
MiniMax is still little-known in the DACH region, but delivers open models with very long context. A look at a provider worth keeping on the radar.
Read more → - Infrastructure2026-05-21
Proxmox VE 9.2: Dynamic Load Balancer for the Cluster
With version 9.2, Proxmox VE gains a dynamic load balancer that automatically distributes running VMs across the cluster; building on the 9.0 generation with Debian 13, new storage features, and SDN fabrics.
Read more → - Security2026-05-20
Patch Management Without SCCM: Updates in SME Daily Operations
Patching is annoying until a vulnerability is exploited. How we automatically keep Linux VMs, Docker containers, and Proxmox hosts up to date, without tools that require their own budget.
Read more → - Strategy2026-05-19
Files & Collaboration: Nextcloud vs. OpenCloud
Nextcloud has dominated the self-hosted market for years. OpenCloud re-enters with Go and oCIS. An honest comparison: when the incumbent, when the challenger is the better choice.
Read more → - AI News2026-05-07
AI Radar: Anthropic's Claude-4 Family (Opus, Sonnet, Haiku)
Anthropic's current Claude generation covers three sizes: Opus for heavy lifting, Sonnet for everyday use, Haiku for speed and cost. What this means for the SME.
Read more → - Strategy2026-05-06
Helpdesk in-house: Zammad as a Ticket System
A helpdesk that turns emails into tickets and does more than Outlook rules. Zammad in-house: setup, SLA management, knowledge base, and the difference from SaaS providers.
Read more → - Strategy2026-05-05
Documentation-as-Code: Wiki.js, BookStack or Markdown in Git?
Internal documentation usually fails at the hurdle of writing. Markdown in Git reduces this hurdle to zero, and an agent that already writes code documents along the way.
Read more → - Infrastructure2026-04-30
Proxmox Backup Server 4.2: S3 Object Storage and Parallel Sync Jobs
Proxmox Backup Server 4.2 brings official S3-compatible object storage as a backend, parallel sync jobs, and server-side encryption of replication; based on Debian 13, Kernel 7.0, and ZFS 2.4.
Read more → - Infrastructure2026-04-22
WireGuard: VPN for Everyday Use
WireGuard is fast, simple, and in the Linux kernel. How we build site-to-site and road-warrior VPNs with it, what to watch for with NAT and MTU, and why it replaces IPsec and OpenVPN in almost every SME scenario.
Read more → - Infrastructure2026-04-21
Monitoring without the baggage: Uptime Kuma in-house
Uptime Kuma brings status monitoring, notifications, and a public status page in one Docker container. What the tool does, how we use it, and where its limits lie.
Read more → - AI News2026-04-10
AI Radar: Google Gemini
Google's current Gemini generation scores with a huge context window and native multimodality, and is deeply integrated into Google Workspace. Strengths and stumbling blocks.
Read more → - Infrastructure2026-04-08
DNS Adblocking Self-Hosted: Pi-hole vs. AdGuard Home
Block ads and tracking network-wide, without browser plugins. Pi-hole (v6) and AdGuard Home compared: setup, DoH/DoT, recursive DNS with unbound, redundancy, and everyday usability.
Read more → - Infrastructure2026-04-07
Git In-House: Gitea and Forgejo
Source code does not belong on foreign servers. Gitea and Forgejo bring GitHub functionality to your own network, CI/CD, Issues, Pull Requests, Actions. Lightweight, open source, ready to deploy in 20 minutes.
Read more → - AI & Development2026-03-20
Codex, Gemini CLI & Co: The Agent Landscape 2026
Claude Code, OpenAI Codex, Google Gemini CLI; the most important agentic coding tools at a glance. Strengths, differences, and why we take a multi-track approach.
Read more → - AI News2026-03-18
AI Radar: GPT-5 from OpenAI
OpenAI's current generation leads in reasoning and multimodality, and remains for many the easiest entry point. What companies should consider.
Read more → - Infrastructure2026-03-17
Docker vs. LXC in Proxmox: When to Use What?
Docker containers or LXC? This question arises for every Proxmox operator. Both have their place, but for different use cases. A practical decision guide, including unprivileged containers and the Docker-in-LXC pitfall.
Read more → - Infrastructure2026-03-03
Reverse Proxy with Caddy: HTTPS by Default
Caddy makes reverse proxying easy; automatic HTTPS, HTTP/3, no certificate headaches, one configuration file. How we secure internal services and make them publicly available.
Read more → - AI News2026-02-21
AI Radar: DeepSeek, From R1 to the V4 Series
DeepSeek has made waves with open weights and strong reasoning. Why the models are technically interesting, and why the hosting location needs careful scrutiny.
Read more → - Security2026-02-20
Intrusion Prevention: CrowdSec instead of fail2ban
fail2ban is yesterday. CrowdSec shares threat data across all participants and blocks attackers before they reach your services. Modern, open source and NIS2-compliant to operate.
Read more → - AI & Development2026-02-12
Claude Code in Practical Use
Anthropic's agentic coding assistant in real project life: What we use Claude Code for, how Skills, MCP servers, and subagents work together, and where the limits lie.
Read more → - Infrastructure2026-02-06
3-2-1 Backup Strategy with Proxmox Backup Server
Backups only matter when they're missing. Proxmox Backup Server, ZFS snapshots, and a Hetzner Storage Box form a 3-2-1-1-0 strategy that is NIS2-compliant and field-proven.
Read more → - Infrastructure2026-01-16
Proxmox HA Cluster: Quorum, Fencing and the End of Single-Host Woes
A single Proxmox host is fine, until it fails. Then everything stops. A three-node HA cluster with Ceph storage is better. What the setup truly requires, why a QDevice solves the two-node problem, and why fencing is no optional detail.
Read more → 
AI & Development2026-01-15Agentic Coding: How AI Agents Transform the Dev Daily Routine
From autocomplete to autonomous agent: What 'agentic coding' really means, where it works today and where human judgment remains indispensable.
Read more →- AI & Development2026-01-01
Email AI: Classification, Routing, and Reply Suggestions Locally
Automatically classify emails, route to the right handler, and generate reply suggestions, with local LLM, n8n, and own mail infrastructure. Productive, not futuristic.
Read more → - Strategy2025-12-18
PDF-Workflows: Automatically ingest incoming invoices into ERP
Incoming invoices via email in, classified into ERP out, without manual data entry. How Paperless-ngx, n8n and a local LLM turn PDFs into structured data and why it must be on-prem.
Read more → - Strategy2025-12-04
Translogica TMS: On-Prem Transport Management
Transportation Management Systems are the backbone of logistics. Translogica on-prem gives us control over routes, shipments, and integrations, without cloud dependency in daily operations.
Read more → - Strategy2025-11-20
Winline ERP In-House: Why We Don't Move Core Software to the Cloud
ERP systems contain a company's most sensitive data. Winline on-prem on our own SQL server gives us full control, over data, customization, and integrations.
Read more → - AI & Development2025-11-06
MCP-Server in-house: Agents with their own knowledge fed
Model Context Protocol gives coding agents tools, but the most valuable MCP servers are those with their own knowledge: internal docs, ERP data, TMS status. So we build them ourselves.
Read more → - Infrastructure2025-10-16
Hosting at Hetzner: Root-Servers and Colocation as the Middle Ground
Cloud is expensive, maintaining your own server room is cumbersome. Hetzner offers root servers and colocation as a middle ground: your own hardware, someone else's power, full control. What makes sense and how we use it.
Read more → 
Storage2025-10-14Storage Journey, Part 3: Arrived at DRBD & LINSTOR
Shared-nothing block-level replication, live-migratable VMs without central SAN and commercial support path. Why DRBD/LINSTOR is the right storage foundation for us.
Read more →- Strategy2025-10-02
Cloud vs. On-Prem: The Honest Calculation 2027
Cloud is not an end in itself and on-premises is not nostalgia. The decision depends on workload, cost structure, and compliance, and it turns out differently in 2027 than it did just five years ago.
Read more → - AI & Development2025-09-16
EU AI Act: Why Local AI Is the Right Answer
The EU AI Act regulates AI applications by risk class, and makes locally operated models the strategically cleanest option. What the law requires and why owning your hardware simplifies compliance.
Read more → - Security2025-09-02
NIS2 and Self-Operation: Compliance through Control
NIS2 obligates more companies to cybersecurity measures than many realize, and places personal responsibility on management. Why self-hosting and self-operation simplify implementation.
Read more → - AI & Development2025-08-19
AI Phone Agents: Voice Bots with Pipecat & Asterisk
Real-time voice pipeline, STT, LLM, TTS; connected to a phone system: Pipecat and Asterisk make it possible. A look at what works today and where the limits lie.
Read more → - Infrastructure2025-08-05
VoIP & Telephony with innovaphone, on-prem instead of Cloud-PBX
Telephony does not need to move to the cloud. Innovaphone brings IP-PBX, SIP trunks, and device management into your own rack. A field report on setup, high availability, and integration.
Read more → - AI & Development2025-07-15
Workflow Automation with n8n, Self-Hosted
n8n connects services, webhooks, and now also AI agents in a self-hosted workflow tool. What sets n8n apart from Zapier and Make, and why we use it for serious automation.
Read more → - Storage2025-07-09
Storage Journey, Part 2: Vitastor, fast, but young
Vitastor promises Ceph-like architecture at significantly lower latency. What we observed in operation, and why 'fast' alone isn't enough.
Read more → - Strategy2025-07-01
Self-Hosted Business Intelligence with Metabase
Dashboards on your own databases, no SQL knowledge needed, no cloud subscription. Metabase brings BI in-house, and makes querying your own data as easy as Googling.
Read more → - AI & Development2025-06-17
RAG Environments Compared: Vector Databases, Chunking, Embeddings
Retrieval-Augmented Generation promises facts from your own documents, but the pipeline behind it decides the quality. Vector databases, chunking strategies, and embedding models in a practical check.
Read more → - AI & Development2025-06-03
LLM Frontends: OpenWebUI vs. LibreChat
Two self-hosted web interfaces that turn local LLMs into a usable chat service. OpenWebUI and LibreChat compared: multi-model, RAG, user management, everyday usability.
Read more → - Strategy2025-05-21
Self-Hosting as a Strategy: Data Sovereignty Instead of Cloud Dependence
Self-hosting is not a hobbyist romance, but an economic and legal decision. When own hardware surpasses the cloud, and when it does not.
Read more → - AI & Development2025-05-20
llama.cpp & GGUF: LLMs locally, efficiently, without a cluster
llama.cpp enables LLM inference on consumer hardware; quantization, CPU/GPU offload, and a file format that needs no cloud budget. The counterpoint to vLLM and when to use which tool.
Read more → - AI & Development2025-05-06
High-Throughput LLM Inference with vLLM on Custom Hardware
vLLM makes LLM inference fast and efficient, PagedAttention, continuous batching, and an OpenAI-compatible API. Here's how we run it on an RTX 5090 and what really matters.
Read more → - Storage2025-04-22
Storage Journey, Part 1: What Ceph Taught Us
Ceph is the Swiss Army knife of distributed storage, powerful, but hungry. Why we started with it, what hardware it really demands, and where it hit limits for our cluster size.
Read more → - Infrastructure2025-04-15
Network Basics: VLANs & Jumbo Frames
802.1Q segmentation and MTU 9000 are not just details for network enthusiasts, but solid tools for more security and throughput. When they help, how Proxmox SDN manages them, and where they cause harm.
Read more → - Infrastructure2025-04-01
Server OS Choice: Ubuntu vs. Alpine
Ubuntu brings stability and support, Alpine minimal image size and musl. When each server OS makes sense, and why the question looks different in the container era than it did five years ago.
Read more → - Infrastructure2025-03-18
SQL Server 2022/2025 Self-Hosted, Not Just on Windows
SQL Server has long run stably on Linux and in containers. What versions 2022 and 2025 bring, how backup and high availability look in self-hosted operation, and what licensing costs.
Read more → - Security2025-03-04
Single Sign-On & IAM: Keycloak Secures Internal Services
One login for all internal services, Keycloak makes it happen without licensing costs. OIDC, SAML, Identity Brokering, and MFA for web applications behind the reverse proxy. Here's how we use it.
Read more → - Security2025-02-21
Mail Security: NoSpamProxy in front of your own mail server
Spam, malware and missing encryption are the three persistent problems of email. NoSpamProxy solves all three as a gateway in front of your own mail server, and brings S/MIME with it.
Read more → - Infrastructure2025-02-18
From VMware to Proxmox: Why We Switched Hypervisors
Licensing costs, uncertainty after Broadcom's acquisition, and the desire for true data sovereignty: why we migrated our virtualization to Proxmox VE; including native ESXi import, VirtIO pitfalls with Windows, and a clean cutover.
Read more → - Infrastructure2025-02-07
Self-Hosted Mail Server: Stalwart & Axigen vs. Exchange
Exchange is expensive and complex. Modern self-hosted alternatives like Stalwart and Axigen offer JMAP, IMAP, and SMTP at a fraction of the complexity. What they deliver, where the differences lie.
Read more → - Innovation2024-12-10
How to Embed AI in Your Company Without Losing Control
Artificial intelligence has become a competitive factor. How to embed AI securely and sustainably in your company, from data foundation to operation.
Read more → - Cloud Solutions2024-12-01
Cloud Migration: The 6-Phase Strategy That Decides Success
From countless migration projects: the 6-phase strategy that decides success or costly failure, from inventory assessment to running operations.
Read more → - Web Development2024-11-15
PHP 8 Performance: Achieve noticeable speed gains
JIT, OPcache and Preloading properly configured; how to turn an average PHP application into a fast one without rewriting the code.
Read more →