DeepSeek ships V4.1-Flash-1017, an MIT-licensed update focused on coding agents

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, a point update to its Flash model rather than a new family. The Hangzhou-based lab frames the release around end-to-end software projects, where an agent plans, writes code, calls tools and iterates until something works.

What changed

The company says the new checkpoint was trained with reinforcement learning using self-feedback, and that it completes tasks faster and with shorter outputs than the previous Flash version, which should lower the cost per task. Its announcement plots DeepSWE v1.1 results against cost per task and describes the update as an improvement on both axes.

The Hugging Face model card lists DeepSeek's own comparison with the earlier Flash. Terminal-Bench 4.0 climbs from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 from 74.2 to 77.0. Humanity's Last Exam without tools moves from 36.8 to 42.6. The card notes that Terminal-Bench was run with the Claude Code harness and DeepSWE with the Mini-SWE-V2 harness, so results depend on the agent scaffolding used. None of these numbers has been independently reproduced in the sources reviewed.

Weights and APIs

Weights are published on Hugging Face and ModelScope under the MIT licence, with a 1M-token context window and inference instructions for vLLM and SGLang. The model card lists a size of 763B parameters but does not state the active parameter count or architecture details, which are left to a separate technical report. DeepSeek recommends a temperature of 1.0 and a generous output budget.

Alongside the model, DeepSeek opened its Search and Fetch APIs, backed by its own search engine, to all API users at $0.75 per 1,000 calls each. Calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free until billing starts on October 31 at 16:00 UTC.

Why it matters

The release keeps pressure on the open-weight segment, where Chinese labs have set the pace on coding benchmarks. A permissive MIT licence lets companies self-host and modify the model without the revenue thresholds found in some rival licences. Teams evaluating it should still test on their own workloads, given that the benchmark gains are vendor-reported and harness-specific.