senn-techAI Radar: Gemini 3.5 Flash & Gemma 4 — Fast in the Cloud, Open at Home
Google is serving both camps at once right now: a fast closed model for the cloud and an open one for your own premises. That very split is what makes the recent releases interesting.
Two Models, Two Worlds
- Gemini 3.5 Flash: optimized for agentic workflows and coding, with much faster output than its predecessors and strong results in tool-oriented benchmarks. Closed, runs in Google's cloud.
- Gemma 4 (12B): the open line, freely available on Hugging Face and therefore self-hostable — the counterpart to the closed Flash variant.
- Speed vs. sovereignty: Flash scores on pace and reach, Gemma on control and self-operation. Both from the same house.
The Sovereignty Fork
This is the heart of the decision, and Google plays it out deliberately in both directions. Gemini 3.5 Flash sits at one end: high performance, direct integration into agentic pipelines, and a very large context window, but every prompt leaves your network and is processed in Google's data centers. Gemma 4 stands at the other end: a smaller twelve-billion-parameter architecture, but with downloadable weights under Google's own permissive Gemma license. Anyone running Gemma keeps inputs on their own hardware: in their own server room, under their own access control, and without any transfer to a third country.
In practice this means: for a quick classification of public data or a coding prototype built in a morning, Flash is often the faster choice. The moment customer data, contracts, or internal knowledge come into play, the calculation flips — and Gemma 4 on your own hardware becomes the cleaner path.
The Caveat
Gemini 3.5 Flash is fast and cheap — but the inputs run through Google's cloud. For personal or confidential data, that remains a GDPR issue. Gemma 4 solves exactly that, at the cost of your own hardware and operations. The choice is not technical but one about data sovereignty.
What Self-Hosting Costs
Gemma 4 is deliberately kept small at twelve billion parameters — that is no accident but the precondition for self-hosting on SME-grade hardware. Quantized to four-bit or FP8, the model can be loaded on a single workstation GPU; the VRAM demand stays manageable. Anyone wanting the full precision advantage needs correspondingly more memory. The price of sovereignty is real but calculable: one GPU, power, maintenance — and the staff to operate the model. Compared with recurring API costs and the permanent dependence on a US provider, that is, for many DACH companies, the more honest calculation.
Our Take
That Google serves both paths in parallel is more honest than it sounds: if you need speed for uncritical tasks, take Flash; if you want sovereignty over your data, host Gemma yourself. We judge it per task — and in the SME world, the answer lands on the open side more often than many expect.
Further Reading
When do you choose Gemini 3.5 Flash over Gemma 4?+
When speed and a vast context window matter and the inputs are uncritical — a classification of public data, a coding prototype built in a morning, or research. Flash runs in Google's cloud and is fast and cheap, yet every prompt leaves your network. The moment customer, contract, or HR data come into play, the calculation flips: then you host Gemma 4 yourself.
Can Gemma 4 be run sensibly on your own hardware?+
Yes, that is exactly the point of the twelve billion parameters. Quantized to four-bit or FP8 the model fits on a single workstation GPU, and VRAM demand stays manageable. The price of data sovereignty is real — one GPU, power, maintenance, and the staff to operate the model — but compared with recurring API costs and dependence on a US provider, that is the more honest calculation for many DACH companies.
What GDPR question does Gemini 3.5 Flash raise?+
The same as with any cloud model: inputs run through Google's data centers, so a third-country transfer takes place. For personal or confidential data that remains a GDPR issue that enterprise contracts mitigate but do not solve. That is precisely where Gemma 4 comes in — anyone who hosts the weights themselves keeps every input on their own network.