Running the open AI model Kimi-K3.1-Pro on a server with 8 GPUs: comparing speed and code optimization ability(zenn.dev)
Writing language: KoreanRead in the original language
Plain summary
- Kimi-K3.1-Pro is a large language model whose weights were released by China's Moonshot AI. Japan's Fixstars installed it directly on its own server and tested its response speed and its ability to quickly fix code.
- On a single server with 8 NVIDIA B300 GPUs, it produced 64.4 tokens per second per user with 64 concurrent sessions. The previous model, Kimi-K3, under the same conditions is 38.7. The measurement is a 4-minute sample, and the official recommendation is at least 30 minutes.
- 10 models solved 12 in-house assembly optimization tasks. Based on the best score among 3 attempts, Kimi-K3.1-Pro ranked 1 with 0.9494 points, and DeepSeek-V4.1-Flash-1017 ranked 2 by a narrow margin.
- The author believes these tasks are an evaluation tailored to the company's own work, so the difference was smaller than the gap in an external composite index (AA Index v5). The license requires separate permission from large API providers, but it states that internal company use is free.
Found on
Zenn ↗ / Published by source
Have you run into something like this? Tell us in a line or two what you did. You do not need to know the product.