DeepSeek V4.1-Flash tops open-weight evals after GA, at a fraction of Kimi K3 cost
DeepSeek
Following yesterday's GA (this is a follow-up, not a repeat): independent evals place the day-old DeepSeek V4.1-Flash at the top of the open-weight class — vals.ai ranks it #1 on the Vals Index (57.86%), narrowly ahead of Kimi K3 at roughly 1/20 of the cost. OpenRouter lists it at $0.15/M input and $0.60/M output tokens with the full 1M-token context. Alongside the weights DeepSeek published the deepseek-recipe inference stack and the DeepSelect and DeepJIT supporting repos; its own numbers beat the larger V4-Pro on Terminal-Bench 2.1, DeepSWE v1.1 and CyberGym.
Why it matters
Independent confirmation that the new CED architecture is competitive at the top of the open-weight class at radically lower cost per token.
Importance: 2/5
Post-release independent evals confirming yesterday's GA release