StudioFlint on RTX PRO 6000 with NVFP4 in vLLM
Original title: Ran Flint on RTX PRO 6000 at NVFP4 in vLLM
The author successfully ran Flint's NVFP4 version via vLLM on an RTX PRO 6000, conducting coding benchmark tests. It scored 82.2 on the LiveCodeBench v6 subset (official: 80.8) and 80 on the SWE-bench Verified subset (official: 74.1). Throughput reached approximately 110-180 tok/s in single-stream mode and peaked at around 2,500 tok/s under 64 concurrent requests. The author considers it a solid daily-use model, weighing whether to adopt it over qwen3.8 27b for its speed gains, and provided docker compose configurations along with a fix for chat_template.jinja.
The author conducted hands-on tests of Flint's coding performance and throughput using vLLM in a custom environment, offering a valuable first-hand reference alongside official data.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.