The model update
DeepSeek has published V4.1-Flash-1017 on Hugging Face under an MIT licence. Its model card calls it a minor update to V4.1-Flash, with gains in reasoning, coding, agentic tasks and visual benchmarks, and says it reduces both task execution time and output length. The context window is one million tokens; DeepSeek recommends a temperature of 1.0, top_p of 1.0 and a maximum output of at least 256,000 tokens.
The company’s figures, which are its own, show the checkpoint ahead of its predecessor on every published benchmark. On Humanity’s Last Exam without tools it scores 42.6 against 36.8. On Terminal-Bench 4.0 it scores 42.3 against 31.2, and on ProgramBench 29.1 against 20.3 — gains of eleven and nine points respectively, which is a large move for a checkpoint described as minor. DeepSWE v1.1 goes from 74.2 to 77.0, NL2Repo-Bench from 65.4 to 68.2, Automation-Bench from 54.8 to 60.5 and Apex Bench from 64.6 to 72.1. The card also compares the model with Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5.
One number does not line up. DeepSeek’s September announcement described V4.1-Flash as a 552-billion-parameter mixture of experts with 8 billion active parameters for input and 16 billion for output. The Hugging Face repository’s tensor metadata for the 1017 checkpoint totals 763 billion parameters across mixed precisions, which is a different accounting rather than a different model, and the company has not reconciled the two publicly.
The swap nobody has to do
The more consequential change is operational. DeepSeek’s API documentation states that deepseek-flash is now served by V4.1-Flash-1017 and that the calling method is unchanged. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired; requests to them route to the new checkpoint and are billed at the Flash price.
For anyone with the endpoint in production, the model underneath changed without a migration step, a version string or an opt-in. That is convenient and it is also the thing that makes reproducing a result from last month hard.
A search engine with a meter on it
Alongside the model, DeepSeek has documented a Search and Fetch API built on its own in-house search engine. Search returns structured results for a query; Fetch retrieves a single page and returns it as Markdown. Both are JSON endpoints under api.deepseek.com with standard bearer authentication.
Each costs $0.75 per 1,000 successful requests, and DeepSeek bills only for HTTP 200 responses that actually return something — failed requests, empty results and timeouts are not charged, and asking for more results with the count parameter does not change the price. Each endpoint allows five requests a second at account level, with independent quotas, and Fetch adds a per-domain limit; going over returns a 429.
The free period is the part with a date on it. Server-side tool calls made through DeepSeek Harness, Claude Code or Codex are free now, and per-request billing begins at midnight Beijing time on 1 November 2026.
What to watch
Two things. Whether independent evaluators reproduce the 1017 gains, since every figure above is DeepSeek’s own. And what happens to usage on 1 November, when a free retrieval layer inside three widely used coding harnesses starts costing money per request.
