DeepSeek V4.1 Flash pushed from broken to 284+ t/s on RTX PRO 6000, +78.7% over vLLM

SergioPaniego · x · 2026-11-20

Developer alvarobartt used Codex (GPT 6.1 Sol) usage credits to spend three days optimizing DeepSeek-V4.1-Flash on NVIDIA RTX PRO 6000 (sm120), going from not even starting on vLLM to 284+ tokens/s on HuggingFace Jobs instances — a 78.7% improvement over vLLM v0.32.0 with DSpark. All changes are open source, with a full article documenting the work.

Original post →

More from coding & agent

coding & agent channel →