DeepSeek V4.1 Flash 1017 is out, with downloadable MIT-licensed weights and improved results across the company’s coding and agent benchmarks. Released on October 17, the update is also live on DeepSeek’s website, app and API.
DeepSeek says the model finishes tasks faster and produces shorter outputs than the previous version. The company describes it as a minor update to V4.1 Flash, although some of its published benchmark gains are substantial.
Coding and agent scores move up
In DeepSeek’s published results, Terminal-Bench 4.0 rises from 31.2 to 42.3. DeepSWE v1.1 moves from 74.2 to 77.0, while ProgramBench increases from 20.3 to 29.1.
| Benchmark | Previous V4.1 Flash | Flash 1017 |
|---|---|---|
| Terminal-Bench 4.0 | 31.2 | 42.3 |
| DeepSWE v1.1 | 74.2 | 77.0 |
| ProgramBench | 20.3 | 29.1 |
These are DeepSeek’s own evaluations. Terminal-Bench used the Claude Code framework, while DeepSWE used Mini-SWE-V2. The scores do not establish how reliably the model will handle every repository or development task.
The release announcement says reinforcement learning with self-feedback was used to improve complete project development. Its demonstrations include a browser racing game and a 3D scene.
Open weights come with a large download
The official model repository provides the weights under the MIT License. It supports image and text inputs, with a listed context window of one million tokens.
The repository’s files total roughly 510GB. That is a download figure, not a statement of the memory needed to run the model. Working memory and conversation context add their own requirements.
The release also omits a Jinja chat template, the formatting instructions many inference tools use for conversations. DeepSeek instead provides reference encoding code and its deepseek-recipe toolkit. Local deployment needs software that handles that prompt format correctly.
Existing API users keep the same model name
API users can continue calling deepseek-flash. DeepSeek says the calling method and pricing are unchanged, with the name now pointing to the October 17 model.
For agent workflows, the company recommends DeepSeek Harness 0.2.1-rc.1 or newer. Readers new to these setups can start with our explanation of how local AI tool calling works.
The weights are available now through Hugging Face and ModelScope, while hosted access avoids downloading the full checkpoint.
