DeepSeek-V4.1-Flash-1017

A post-training update to DeepSeek-V4.1-Flash, released October 17, 2026 with open weights under MIT on HuggingFace and ModelScope. Its config.json is identical to the September checkpoint, so the architecture is unchanged: a 552B MoE with 16B active at decode, the Causal Encoder-Decoder layout, CSA2 attention, the Engram memory table and a 1M-token context. DeepSeek says it trained the update with "reinforcement learning with self-feedback" for end-to-end project building, and claims "a Pareto improvement in performance and cost": higher scores, faster task completion and shorter outputs. It now serves the deepseek-flash API name at the same Flash price, and it is the default checkpoint of DeepSeek's DeepInfer inference framework.

DeepSeek's self-reported gains over the September checkpoint: Terminal-Bench 4.0 42.3 vs 31.2, HLE without tools 42.6 vs 36.8, ProgramBench 29.1 vs 20.3, Apex Bench v1.1 72.1 vs 64.6, Automation-Bench 60.5 vs 54.8, Agents' Last Exam 35.4 vs 31.8, DeepSWE v1.1 77.0 vs 74.2 and NL2Repo 68.2 vs 65.4.

The independent number moves the same way. Artificial Analysis scored it 45 on Intelligence Index v4.3.2, six points above the original, and marked the September entry deprecated in its favour. On v5.0 it scored 32 at max effort, the highest of any model with released open weights at v5's launch. On v5.0.1 (November's CritPt v2 regrade) it reads 33 (32.6 unrounded), ahead of GLM-5.3 (32, 31.7 unrounded), MiMo-V2.6-Pro (30) and Ling-3.1-flash-1019 (30), and well above V4-Pro (23). Kimi K3.1 Pro (46) has since overtaken it among open-weight models, after its weights were released on November 2. Its strengths on AA's runs are AutomationBench-AA (66.6%), AA-Briefcase (Elo 1570) and FrontierCode 2.0 (28.5%, the best open-weights score). Its weak spot is calibration: a 35.9% AA-Omniscience non-hallucination rate. It is also cheap to run, at $0.60 per index task on $0.30 / $1.20 per million tokens.