DeepSeek updates Flash model and open weights
DeepSeek released V4.1-Flash-1017 with open weights and higher developer-reported coding scores. The API name and pricing stay unchanged.
DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, updating its website, app and API. It also published the model weights under the MIT license.
In the developer's evaluations, the DeepSWE v1.1 software repair score rose from 74.2 to 77.0. Terminal-Bench 4.0, which tests work in a terminal, rose from 31.2 to 42.3. DeepSeek used Mini-SWE-V2 for the former and Claude Code for the latter.
DeepSeek also reports shorter task runtimes and outputs, without specifying percentage reductions in the announcement. These are developer-reported results, not evidence of the same gains across every workflow.
API users can continue to call deepseek-flash with the same request method and pricing. DeepSeek recommends version 0.2.1-rc.1 or later of its Harness tool.
Sources
- [1]
- [2]
Images: generated with GPT Image 2.5 (illustrative).