Olares One first look: AI mini PC with 24 GB VRAM, not 32

The Olares One is a mini PC for local AI that Olares offered on Kickstarter in December 2025. In Caschy's first look of 24 October 2026 he is "torn". Only the retail launch is new, since the hardware has been known since November 2025. It still shows well what 8 GB less video memory means for local language models.

Every number in the text says where it comes from. Verified means we read at least two independent sources. Inferred is our own calculation. Vendor claim is a value nobody has checked independently.

The graphics chip is the laptop variant of the RTX 5090 with 24 GB. Confirmed by Caschy (24 October 2026) and Notebookcheck (9 December 2025). (Quelle: Simple Icons, file nvidia.svg (CC0))

What is inside

The graphics card is an RTX 5090 Mobile with 24 GB GDDR7. Caschy writes: "This is the notebook variant, not a desktop card with 32 GB." Notebookcheck confirms it (verified). Nvidia specifies 95 to 150 watts of power for the laptop card, and the manufacturers set the value themselves. Olares lists 95 watts in silent mode and 175 watts in performance mode in its shop (vendor claim). According to Nvidia, the desktop card has 32 GB on a 512-bit bus and 575 watts.

The rest according to the shop (vendor claim, Caschy also confirms RAM and SSD): Core Ultra 9 275HX with 24 cores, 96 GB DDR5-5600, 2 TB NVMe, Thunderbolt 5, Wi-Fi 7, a 330-watt power supply, 2.15 kg. On software: Olares OS is open source under AGPL-3.0, builds on Kubernetes and, according to the GitHub project, can also be installed by script on Ubuntu 22.04 to 25.04 or Debian 12/13. Caschy describes setup through the LarePass app and Tailscale. Whether Windows runs directly is contradictory: the shop mentions a Windows VM over RDP, Notebookcheck says Windows or Linux as the operating system. That is unresolved.

DB Tech reviews the Olares software and the hardware of the One. The video is from 14 November 2025, during the Kickstarter phase, and links the campaign in its description.

Video memory decides which models run

With local models, VRAM limits what starts at all. In November 2025 Notebookcheck wrote that buyers would have to "ignore larger models completely" or settle for quantized versions. We checked what that means for a model we run ourselves.

Qwen3.8-27B in the IQ4_XS quantization is 17.03 GB, which is 15.9 GiB. With vLLM and FP8, our post Qwen3.8 on the RTX 5090 measured 19.41 GiB for the weights and 6.47 GiB for 184,320 tokens of context cache, which is 36.8 KiB per token. Using that value and an assumed overhead of 1.5 GiB for CUDA and display (not measured), a context of 128,000 tokens gives:

Our own calculation (inferred): weights from our measurements, KV at 36.8 KiB per token from the FP8 measurement also applied to the GGUF, overhead of 1.5 GiB assumed. (Quelle: Own measurement and calculation, October 2026)

The 4-bit variant fits under 24 GB at 21.9 GiB, the FP8 variant at 25.4 GiB does not. If you stay with FP8, the 3.1 GiB left after the weights and overhead cover about 85,000 tokens. With the 4-bit file it is roughly 165,000 to 187,000 tokens, if you leave out the vision projector. On a 32 GB card we measured the same model with 262,144 tokens of context: 120 to 146 tokens per second in a single stream with speculative decoding at depth 4 (speed post of 19 October).

As a rule of thumb the vendor gives 0.5 bytes per parameter at 4-bit plus 20 to 30 percent for context (vendor claim). On 24 GiB that comes to about 38 billion dense parameters, on 32 GiB about 51 billion. An MoE model like Flash-Next with 125 billion parameters needs 90.9 GiB in the smallest quantization we have measured (series of 29 August 2026) and fits on neither card.

Speed: what the vendor measures and what bandwidth allows

In its blog post of 4 November 2025, Olares published figures (vendor claim, unchecked by third parties). Qwen3-30B-A3B in AWQ 4-bit reaches 157.17 tokens per second on vLLM with one request and 81.26 with eight. Gemma3-27B gets 38.02 and 28.81, GPT-OSS-20B gets 139.00 and 72.20.

Vendor claim, not independently checked. The first three models ran on vLLM, GPT-OSS-120B on llama.cpp. (Quelle: Olares blog, 4 November 2025)

One outlier matters. GPT-OSS-120B does not fit in 24 GB, so layers move to system memory, and according to the vendor llama.cpp drops from 36.16 to 4.44 tokens per second when eight requests run at once. Offloading to the 96 GB of RAM is therefore a cliff that costs most of the speed.

An upper limit for a dense model follows from memory bandwidth. The laptop card has a 256-bit bus. At 28 Gbit/s per pin that is 896 GB/s (inferred). Every generated token requires reading all 17.03 GB of weights once, which yields at most about 53 tokens per second without speculative decoding (inferred). The vendor's figure for Gemma3-27B, around 38, is in the same range. A dense 27B model only gets past 53 with speculation or a smaller quantization. Sparse models with about 3 billion active parameters, like Qwen3-30B-A3B, fit in about 15 GB at 4-bit and are the fastest there for that reason.

Price and terms

Four price points sit side by side, from different dates and currencies: 2,490 euros as the Kickstarter price according to Notebookcheck of 9 December 2025, 2,999 US dollars according to Notebookcheck and the Olares blog of November 2025, 4,999 US dollars in the shop today, and "around 4,550 euros" from Caschy on 24 October 2026. Since the campaign the price has almost doubled, and it should be checked on the day of purchase.

The shop says "Ships within 24 Hours" and the shipping policy says two working days. The warranty is two years from delivery (vendor claim). The terms state that sales are final, and next to that sit the 14-day EU right of withdrawal and a seven-day window for reporting devices that arrive defective. These do not fit together, and we have not resolved it. The noise level (19 dB idle, under 39 dB at full load) is also a vendor claim.

What you can check before buying

Calculation path from this post. The per-token value is in the vLLM or llama.cpp startup log. (Quelle: senn-tech, own calculation)

If you plan around a specific model, add the weight file size, the context cache and about 1.5 GiB of overhead, and compare the sum with 24 GiB. The vLLM startup log shows the cache per token in the "GPU KV cache size" line, and llama.cpp prints it in the startup log of your own model. Measure speed yourself, for example with llama-bench -m file.gguf -p 512 -n 128, and skip the vendor table. If a target model lands at about 10 tokens per second or less after offloading, it is no use as a working model. The return and withdrawal rules can be confirmed in writing with the vendor before ordering. Because Olares OS can be installed on Ubuntu and Debian, the software stays usable without the box.

Caschy's tests with measured values are still to come. Until then the Olares One is a device whose hardware is well documented and whose speed has only been measured by the vendor.