DeepSeek Updates V4.1 Flash to 1017 With Open Weights and Web Search APIs

DeepSeek kept the product name and swapped what sits behind it. According to the company, V4.1-Flash-1017 does more on agent and coding work while cutting output length and runtime.

AI Desk · Published October 17, 2026
First look: this report is based on DeepSeek’s official release notes, pricing page, and Hugging Face model card. Maitz has not run the model hands-on.

What shipped?

According to DeepSeek, V4.1-Flash-1017 is a minor update to DeepSeek-V4.1-Flash with substantial gains in reasoning, coding, agent tasks, and visual benchmarks. The company also reports shorter outputs and faster task completion. It describes the change as a Pareto improvement on the performance–cost curve.

The build is live on the website, the app, and the API. The API still uses the model name deepseek-flash. Legacy names such as deepseek-v4-flash are still accepted, but the docs say those models have been retired and requests are routed to 1017 at Flash pricing.

DeepSeek also highlights reinforcement learning with self-feedback for end-to-end project building. Official demos include a browser racing game, a Portal 2-style recreation in Godot, and 3D scenes. Those are company examples, not Maitz tests.

What do the company numbers say?

On the Hugging Face model card, DeepSeek publishes a comparison table against the previous Flash build and other models. These figures belong to the company. We did not re-run them.

The same table also lists GPT-5.6 Sol and Claude Opus 5. On DeepSWE, DeepSeek reports 77.0 versus 73.0 and 74.0. On Terminal-Bench 4.0 its score sits below Opus 5. Treat those as company claims, not an independent Maitz eval.

Price and access

According to the official pricing page, Flash rates did not change with 1017. Off-peak: $0.15 per million input tokens on a cache miss, $0.003 on a cache hit, and $0.6 per million output tokens. Peak rates are double. Context length is 1M tokens, and the listed maximum output is 384K.

The weights themselves are downloadable under MIT. For teams that self-host, that means the files can be used under MIT terms without a separate model commercial license. Exact local VRAM needs are not spelled out on the short card; the Hugging Face repo points to an inference folder.

Search and Fetch: live web access

On the same day, DeepSeek opened two tools to all API users. Search finds web pages for a query and returns structured results. Fetch pulls a single page and returns Markdown. According to the company, both run on its in-house search engine.

Price: $0.75 per 1,000 successful requests on each endpoint. Until October 31, 2026 at 16:00 UTC, Search and Fetch calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex are free for DeepSeek API users. Billing starts after that.