DeepSeek ships V4.1-Flash-1017 open weights and APIs

DeepSeek's Oct. 17 V4.1-Flash-1017 is a 552B MoE scoring 77.0 on DeepSWE, with Search APIs and open weights.

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, posting open weights under deepseek-ai on Hugging Face and, per The Decoder, also on ModelScope, with web, app, and API access. OpenRouter describes the deepseek-flash snapshot as a 552B sparse mixture-of-experts model, the first on DeepSeek's Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output, with native image understanding, a 1 million-token context, and compressed KV caching aimed at coding, terminal, and computer-use agents. Listed pricing is $0.15 per million input tokens and $0.60 per million output tokens, and DeepSeek says it exceeds V4 Pro on performance, speed, and task completion time while scoring higher, running faster, and producing shorter outputs than V4.1-Flash. MarkTechPost reports a DeepSWE score of 77.0; The Decoder gives 77.0% at just under $0.20 per task, or about 72% near $0.10, versus 74.2% at roughly $0.50 for the predecessor, and says DeepSeek claims wins over Claude Opus 5, GPT-5.6-Sol, Kimi K3, and GLM-5.3, though independent benchmarks are not yet available. The company also opened Search and Fetch APIs at $0.75 per 1,000 calls, free in some tools until October 31, 2026. Early Hacker News coverage had 45 points and 14 comments but no specs; sources agree on the 77.0 DeepSWE figure, with later reports adding cost and comparison detail rather than contradicting it.

• DeepSeek released the October 17, 2026 snapshot DeepSeek-V4.1-Flash-1017, with open weights on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope plus web, app, and API access.
• OpenRouter lists it on the deepseek-flash API as a 552B sparse MoE, the first on DeepSeek's Causal Encoder-Decoder architecture, activating 8B parameters on input and 16B on output, with native image understanding, 1M context, and…
• Listed token pricing is $0.15 per million input tokens and $0.60 per million output tokens; DeepSeek says it exceeds V4 Pro on performance, speed, and task completion time, and beats V4.1-Flash on scores, speed, and output length.
• DeepSWE is reported at 77.0 (MarkTechPost) and 77.0% at just under $0.20 per task, or about 72% near $0.10, versus 74.2% at roughly $0.50 for the predecessor (The Decoder).
• DeepSeek claims it beats larger models including Claude Opus 5, GPT-5.6-Sol, Kimi K3, and GLM-5.3 on several coding and agent tests; The Decoder says independent benchmarks are not yet available.
• Search and Fetch APIs cost $0.75 per 1,000 calls and are free in some tools until October 31, 2026; post-training uses reinforcement learning with self-feedback, without human evaluators or external reward models.
• A Hacker News item had 45 points and 14 comments and contained no model-card text, size, architecture, or benchmarks.