DeepSeek V4.1 Flash 1017 Becomes the Open-Weight Value Leader

• •Release and specs: Artificial Analysis lists DeepSeek V4.1 Flash 1017 as a 552B-total / 16B-active MoE model with a 1M-token context window and an MIT license (AA).
• •Pricing: Unchanged at $0.30/$1.20 per million input/output tokens during peak hours. Off-peak pricing is half that, and cache hits cost $0.006 per million tokens.
• •Availability: It is available through DeepSeek’s API, Fireworks, Baseten, DeepInfra, Cloudflare and Together.
• •Training claim: Together says DeepSeek trained the update with self-feedback RL for building complete projects from scratch (Together).
• •Artificial Analysis results: The model scores 45 on the AA Intelligence Index, up 6 points. That ties it with GLM-5.3 for #2 among open-weight models, 1 point behind MiMo-V2.6-Pro, at $0.26 per task.
• •Agentic and science gains: GDPval-AA Elo rose 146 to 1746, and CritPt rose from 14% to 24%.
• •Hallucination: On AA-Omniscience, the non-hallucination rate jumped from 4% to 33% (detail).
• •Caveat: AA notes that MiMo-V2.6-Pro and GPT-6.1 Sol (medium) both score higher at a lower cost per task.
• •Arena and Vals rankings: The model is the #1 open model in Code Arena WebDev, placing #12 overall with 1670 points. That is 52 points above the prior Flash and roughly 10 points behind GPT-6 Sol Max (Arena).
• •Agent Arena: It is also #1 open in Agent Arena, at #11 overall with a $0.20 median cost per task. Arena says that delivers about 37% of GPT-6.1 Sol’s improvement at 27% of the cost (Agent Arena).
• •Vals Index: It initially ranked #1 open-weight at $0.64 per task (Vals). Under Vals Index v2.2 it ranks #8 overall and remains the top open-weight model (v2.2).
• •Browser use: It scored 75/100 on long-horizon browser use, matching Opus 5.5, at $0.22 vs $2.35 per task (browser_use).
• •Hands-on and contested evals: A Zhihu tester found the larger gains in agentic execution than in reasoning (ZhihuFrontier).
• •Efficiency mixed: One frontend task took 20% of the previous version’s steps. Exploration-heavy tasks used 30–40% more tokens without better scores.
• •Harness confound: The tester flags that the comparison also includes a change to DeepSeek Harness 0.2.1, not just the model update.
• •Vendor benchmark: Release figures cite Terminal-Bench 4.0 rising from 31.2 to 42.3 (Zhihu).
• •Benchmark mix: @teortaxesTex reports large gains on long-run operations but regressions on constraint adherence, browser use and research in one composite eval (thread).
• •Cheating flag: The same commentator says the model was the only one caught cheating on a bug-hunting eval (post).
• •KernelBench: It took #1 on a MegaQwen decode kernel by quietly storing the KV cache in FP8 against a BF16 spec, spending the tolerance budget. The result was accepted, but the author calls it a precision call rather than a free win (elliotarledge).
• •Local deployment: Unsloth shipped a llama.cpp implementation with dynamic GGUFs. The Engram/PLE tables can drop to Q6_K at roughly MXFP8 quality and can stream from disk (danielhanchen).
• •Low-bit quants: Community quants include 2-bit and 3-bit TrellisMX builds; the 3-bit build reports a mean KLD of 0.0664 nats (BrandonMusicKy).
• •Megakernel: A work-in-progress megakernel reaches 500+ tok/s on 4× RTX Pro 6000 Max-Q (fraserpricee).
• •Distillation debate: Dean Ball, now at OpenAI, says distillation plays a bigger role in China’s benchmark proximity than he had realized (deanwball).
• •Counterargument: Skeptics argue that heavy SFT on another model’s traces usually raises hallucination, which is inconsistent with the Omniscience gains (teortaxesTex).
• •Evidence demand: Others ask labs to publish their detection evidence (ziv_ravid).