DeepSeek updates Flash AI model, claiming faster coding tasks and shorter outputs
DeepSeek released V4.1-Flash-1017 on 17 October, an update available through its website, app and API, with open weights under the MIT licence. The company says it improves benchmark scores, completes tasks faster and produces shorter outputs; API pricing is unchanged. DeepSeek’s results are company-run and have not been independently reproduced, according to The Decoder and DEV Community.
Bottom line — The deepseek-flash API now serves the new checkpoint, while independent testing of its benchmark gains is still pending.
Go deeper 6
-
DeepSeek’s figures show Terminal-Bench 4.0 rising from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1.
-
The Decoder reports a DeepSWE score of 77.0 at an average cost of just under €0.17 per task (about $0.20), compared with 74.2 at about €0.43 (about $0.50) for the previous version; these figures come from DeepSeek’s chart.
-
Comparisons with rival models are mixed: MindStudio says Opus 5 leads on several reasoning and vision benchmarks, while DeepSeek’s model card shows the new Flash ahead on some coding and agentic tests.
-
DeepSeek says the update uses reinforcement learning with self-feedback to improve end-to-end project development, including game creation and 3D modelling.
-
The MIT-licensed weights are available on Hugging Face and ModelScope; MindStudio notes the release does not include a standard Jinja chat template.
-
DeepSeek opened its Search and Fetch APIs to developers at €0.69 per 1,000 successful requests (about $0.75); calls through supported server-side tools are free until 31 October 2026 at 16:00 UTC.