DeepSeek ships a faster, cheaper Flash

DeepSeek announced on X this morning that DeepSeek-V4.1-Flash-1017 is live on the company’s website, app, and API. It is served under the existing deepseek-flash API name — same name, same price — and the old deepseek-v4-flash and deepseek-v4-flash-vision-exp names still work, served by the new checkpoint. It is an updated post-training checkpoint of the V4.1-Flash that launched September 10, and DeepSeek calls the whole thing “a Pareto improvement in performance and cost”: better scores, faster tasks, shorter outputs.

Under the hood it is the Causal Encoder-Decoder (CED) sparse mixture-of-experts design: a 552B-parameter backbone plus 196B Engram parameters, activating 8B parameters per token on input and 16B on output, with a 1M-token context window and up to 384K output tokens. The FP4 KV cache lands at 890 bytes per token — about a quarter of the previous Flash generation — and image understanding is native to the architecture rather than bolted on. A companion technical report, “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression,” puts the memory story front and center.

It is trained with reinforcement learning with self-feedback for end-to-end project work — DeepSeek’s examples run from game creation to 3D modeling — and tuned for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. DeepSeek recommends Harness v0.2.1-rc.1 or later for the best experience with the new checkpoint.

The headline score is 77.0 on DeepSWE v1.1, up from 74.2 — ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on that coding benchmark in DeepSeek’s own table. Terminal-Bench 4.0 jumps to 42.3 from 31.2, past GPT-5.6 Sol’s 39.9 in the same table but still behind Opus 5 at 51.8. DeepSeek says the new checkpoint beats the previous Flash on all 11 benchmarks it reported, by 2.0 to 11.1 points — including Automation-Bench (60.5), Agents’ Last Exam (35.4), and Apex Bench (72.1).

The honest footnote: these are DeepSeek’s figures, and no independent run has been published yet. And “Pareto improvement” is DeepSeek judging its own product. On raw reasoning Opus 5 still leads — HLE without tools is 42.6 for Flash-1017 versus 56.3. If you run coding agents on DeepSeek’s stack, test the new checkpoint in your own harness before moving workloads; launch-day tables tell you what the vendor measured, not what you’ll get.

The second half of the announcement matters as much as the model. DeepSeek opened its Search and Fetch APIs to every developer. Search returns ranked web results for a query — 1 to 50 results, with site:, lang:, after:, and before: operators, rerank scores, and authority tiers. Fetch returns a single page as Markdown. The endpoints are POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com, and DeepSeek says its in-house engine serves hundreds of millions of requests daily.

Pricing is $0.75 per 1,000 successful requests per endpoint, with empty results and failures free and a 5 QPS cap per account. Server-side calls routed through DeepSeek Harness, Claude Code, or Codex stay free until billing starts at the end of the month — DeepSeek’s docs say October 31 at 16:00 UTC, one write-up says November 1, so treat October 31 as the safe line. The practical read: a cheaper coding model plus a first-party web layer behind one API key, which is exactly the stack an agent builder wants.

Look at what the model was tuned for: DeepSeek Harness, Tencent CodeBuddy, WorkBuddy — tools DeepSeek either built or has a direct relationship with through Tencent. The original September 10 launch did the same thing: model update, harness version bump, named partner access, all in one announcement. Twice in six weeks is a strategy, not a launch: DeepSeek is pulling developers toward a stack it built end to end.

Every major lab ties its best performance to its own interface, so this is not underhanded. What is notable is how openly DeepSeek is doing it while keeping the open-weights reputation: the weights went up on Hugging Face and ModelScope the same day, MIT-licensed, and the harness is MIT too — nothing stops a developer from running V4.1-Flash-1017 through a competing tool instead. One more thing to watch: DeepSeek has now named V4.1 Pro twice, and the Flash line keeps outpacing it. The next flagship is coming.