
Three Used V100 GPUs Run Qwen 3.6 35B at 100 Tokens a Second
A homelab builder hit 100.8 tokens per second on Qwen 3.6 35B with three used Nvidia V100 GPUs in a rig priced at about $1,042. Here is the parts list.
Used V100 GPUs keep proving that a capable local LLM box does not need new hardware. In a benchmark write-up published November 17, the builder behind the esologic blog ran Qwen 3.6 35B at 100.8 tokens per second on three decommissioned Nvidia Tesla V100 16GB cards, in a self-hosted rig whose core parts came to about $1,042. Tom's Hardware picked up the results on November 19.
- Speed: 100.8 tokens per second at 4K context and 72.7 at 128K, running Qwen 3.6 35B at Q8 in llama.cpp.
- Core parts cost: $655.16 for three V100s, $211 for the board, $147.70 for the power supply and $28.20 for one CPU.
- Not included: RAM, storage, case and cooling, plus the second CPU, which lifts the total to about $1,070.
- Efficiency: roughly 0.02 kWh of GPU energy for a 128K-context run, about half what three P100s needed.
What Hardware Is in the $1,042 Local LLM Rig?
Every build in the test used the same platform: a Supermicro X10DRG-Q dual-socket board, two Intel Xeon E5-2687W 12-core CPUs and a Corsair HX1500i power supply. The builder then swapped in one to three Tesla K80, P100 or V100 16GB cards for nine configurations and more than 130 LLM tests covering Qwen 2.5, Gemma 4 and Qwen 3.6 at several quantization levels and context sizes.
Why three cards? Qwen 3.6 35B at 8-bit quantization does not fit on one or two V100s. The builder notes the listed prices will not perfectly match what you find on eBay today.
There is a catch that keeps these cards cheap. Data center GPUs like the V100 run hot and ship without fans, so each one needs forced air. The builder 3D printed a blower duct per card and says a standalone cooler is close to release.
How Fast Are V100 GPUs for Local AI?
The builder set 70 tokens per second as the bar for real-time interaction. Three V100s cleared it at every context length tested on Qwen 3.6 35B, sliding from 100.8 tokens per second at 4K to 89.4 at 32K, 83.5 at 64K and 72.7 at 128K.
Model architecture matters as much as the silicon. Qwen 3.6 35B is a mixture-of-experts model that activates only about 3 billion of its 35 billion parameters per token. The dense Qwen 3.6 27B managed just 24.3 tokens per second on the same three cards, about a quarter of the speed.
A few other data points for planning a build:
- Two V100s at Q4: 105.4 tokens per second at 4K and 77.1 at 128K, so 4-bit quantization gets the same model running on a cheaper two-card setup.
- Three P100s at Q8: 55.2 down to 40.2 tokens per second, below the real-time bar.
- More cards add memory, not speed: extra GPUs let you load bigger models rather than making a given model faster.
Tom's Hardware lined the results up against published figures for an M5 Max Mac Studio and found the V100 box slightly behind at short contexts and ahead at 64K and 128K, with the caveat that the machines used different runtimes.
Should You Build a Used GPU Server?
If you are comfortable with server boards, loud fans and DIY cooling, the value is hard to ignore: a 35-billion-parameter model at conversational speed for roughly the price of a midrange graphics card. The trade-offs are power draw, noise and aging hardware with no warranty. For a larger take on the same idea, see how a lawyer built a private AI server from eight V100s, and note that the stable llama.cpp v1.0.0 API makes multi-GPU setups like this easier to maintain. More self-hosted builds live in our mini computer coverage.
Sources: esologic — November 17, 2026; Tom's Hardware — November 19, 2026.
More Mini Computers Stories

GUITION ESP32-S31 Kit: 7-Inch Touch Display and Ethernet
The GUITION JC1060F570C pairs an ESP32-S31 with a 7-inch 1024x600 touchscreen, Ethernet and Wi-Fi 6 for about $29, just as Arduino core 4.0.0 adds support.

ANAVI I2C TPM 2.0: A $20 Security Chip for Any Raspberry Pi
The ANAVI I2C TPM 2.0 adds hardware key storage and disk encryption to any Raspberry Pi for $20, with a driver already in Raspberry Pi OS. How it works.

Flipper One Kickstarter: $399 Linux Multi-Tool Ships 2027
Flipper One is now on Kickstarter at $399, with an RK3576, 8GB RAM, Wi-Fi 6E and LoRa, 5G and SDR add-ons. First units are estimated for July 2027.
