DeepSeek swaps in V4.1-Flash-1017 and points harness users to 0.2.1

DeepSeek has swapped the checkpoint behind deepseek-flash. Its API changelog, dated 2026-10-17, says V4.1 Flash is now DeepSeek-V4.1-Flash-1017 on the web, the app, and the API. Callers still pass deepseek-flash. The calling method and pricing stay the same.

That is the part easy to miss. Jobs pinned to deepseek-flash since the September 10 V4.1-Flash release are now hitting a different model, with different agent scores and a new harness recommendation, and the model string did not change. SiliconANGLE, in an October 18 recap, identified the company as Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. and reported the same model ID, unchanged prices, and the same harness note.

DeepSeek describes the update as a performance-cost Pareto improvement over the previous V4.1 Flash. The changelog says task runtime and output length are down, so users can finish work at higher speed and lower cost. It does not say by how much. The Hugging Face card calls 1017 a minor update that reduces task execution time and output length. SiliconANGLE noted that DeepSeek did not quantify the reduction. The API rate is not what moved. Pricing is unchanged; the cost claim is about shorter, faster runs.

Scores on the October 17 changelog match the model card:

HLE without tools: 42.6, and 44.0 on the text-only subset, against 36.8 and 39.1 on September 10.
Terminal-Bench 4.0: 42.3, from 31.2.
DeepSWE v1.1: 77.0, from 74.2.
ProgramBench: 29.1, from 20.3.
NL2Repo-Bench: 68.2, from 65.4.
Automation-Bench v1.0.6: 60.5. The September changelog listed Automation-Bench at 54.8 and did not use the v1.0.6 label.
Agents' Last Exam: 35.4, from 31.8.
Apex Bench v1.1: 72.1. The September changelog omitted this row. The model card lists 64.6 for the prior V4.1-Flash.
Chartography with tools: 83.0, from 78.9.
BabyVision with tools: 92.5, from 89.6.
ZeroBench-main with tools: 51.0, from 49.0.

The footnotes matter. Terminal-Bench 4.0 was tested with Claude Code, and DeepSWE with Mini-SWE-V2. The model card states the same harness split. ZeroBench-main was evaluated with Pass@5. The HLE figure in parentheses is the pure-text subset only. These are DeepSeek's published numbers, not an outside rerun.

The other change under the same name is the harness version. DeepSeek says it optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy, and other platforms, and tells users to upgrade to DeepSeek Harness 0.2.1-rc.1 or later for the best experience with the new model. SiliconANGLE describes CodeBuddy as Tencent's coding assistant and WorkBuddy as a desktop agent for office work that has offered V4.1-Flash since September 10. The model card licenses the repository and the weights under MIT, which SiliconANGLE also reported.

The same October 17 entry opens DeepSeek's in-house Search API for web search and fetch. DeepSeek says the engine handles hundreds of millions of search requests a day, indexes hundreds of billions of pages, supports multilingual search, and filters SEO spam, GEO manipulation, and low-quality generated content. The changelog points to a Search and Fetch guide for usage details.

On the card, 1017 is a minor update to V4.1-Flash. The September 10 notes called that model the smallest in the new architecture family, with native multimodal visual understanding. The card lists 763 billion parameters. Recommended sampling is temperature 1.0 and top_p 1.0, a 1 million token context window, and max tokens of at least 256K. There is no Jinja chat template. A Python reference in the encoding folder covers multi-turn chat, tool calling, thinking mode, and images. deepseek-recipe, Rust libraries with Python bindings, turns Chat Completions and Responses API requests into V4 and V4.1 prompts. The card says no Hugging Face Inference Provider is serving the model.

If you already call deepseek-flash, treat the September scores as stale and rerun the agent tasks you care about on this checkpoint. Move to DeepSeek Harness 0.2.1-rc.1 or later before you judge the update on an older harness, and read the October 17 changelog and the Hugging Face card for the evaluation footnotes.