DeepSeek V4.1 Flash 1017: Same Price, Higher Scores, MIT Weights

DeepSeek V4.1 Flash got a quiet but meaningful upgrade on Saturday: the 1017 build now serves the default deepseek-flash API model at the same price, posts higher scores on DeepSeek's own benchmark table, and ships open weights under MIT. The lab also opened its Search and Fetch APIs to everyone at $0.75 per 1,000 calls.

What actually happened

DeepSeek published a news post titled "DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real" on October 17. The short version: a minor version bump to V4.1 Flash that the company says improves reasoning, coding, agentic tasks, and visual benchmarks while cutting task time and output length. Investing.com, via Yahoo, reported the same release and the Search API news the same day.

The update is optimized for DeepSeek's own Harness, plus Tencent's CodeBuddy and WorkBuddy. DeepSeek recommends Harness v0.2.1-rc.1 or later for the best results. The headline pitch is end-to-end project building, with demos including a 3D Forbidden City model, a web racing game, and a Portal 2 recreation in Godot. Those are company showcase examples, not independent tests.

On the business side, the Search and Fetch endpoints, powered by DeepSeek's in-house search engine, are now open to all API users at $0.75 per 1,000 calls per endpoint. Server-side tool calls through Harness, Claude Code, and Codex are free until billing starts on October 31 at 16:00 UTC.

DeepSeek V4.1 Flash benchmarks, read carefully

The Hugging Face table compares the 1017 build with the previous V4.1 Flash and with outside models. Selected jumps: HLE without tools 42.6 versus 36.8, Terminal-Bench 4.0 42.3 versus 31.2, ProgramBench 29.1 versus 20.3, and Automation-Bench 60.5 versus 54.8. DeepSWE v1.1 moved from 74.2 to 77.0, which Investing.com also cited.

Two caveats. These are vendor-run numbers with vendor-chosen harnesses (Claude Code for Terminal-Bench, Mini-SWE-V2 for DeepSWE). And the same table shows the update still trailing the strongest closed model listed on several rows. A real upgrade, yes; a crown, not by its own chart.

Same price, better scores, open weights. The boring version of an AI launch is often the one developers actually use.

What Polymarket thinks

There is no market on this exact release, but there is a related one: "Next DeepSeek Pro released by...?" At ~06:02 UTC the October 31 rung traded around 10% Yes and November 30 around 19.5%. A Flash refresh is not a Pro model, so this release does not resolve that market. The book is thin, with roughly $25K lifetime volume and no trades in the prior 24 hours, so read it as a vibe check. The live board is on the Polymarket DeepSeek Pro event.

For the open-weight race this week, compare with Muse Spark Open from Meta and Reflection Beam. Both landed in the same seven-day window.

Why a Flash refresh matters

Flash is DeepSeek's cheaper, faster tier, and it is the one most third-party tools wire in by default. That makes a same-price upgrade more important in practice than a flashy flagship launch. Coding agents, IDE plugins, and hobby projects that never touch a settings page quietly get the new behavior overnight.

The shorter-output claim also matters for cost. If the model really finishes tasks with fewer tokens, the effective price per completed job falls even though the per-token price is unchanged. That is the "Pareto" framing in DeepSeek's post: more capability without more spend. It is a claim worth testing on your own prompts rather than taking on faith.

What to do as a reader

If you already call deepseek-flash, you are on the new build: the API name, calling method, and Flash pricing did not change, and legacy names route to the 1017 model. Re-run your own evals before assuming the vendor chart applies to your workload, especially if you rely on output length for cost estimates.

If you want to self-host DeepSeek V4.1 Flash, note the release ships without a Jinja chat template; DeepSeek points to a reference encoder and its deepseek-recipe libraries instead. Budget for the Search and Fetch billing switch on October 31. Nothing here is investment advice.

Sources: DeepSeek news post, Hugging Face model card, and Investing.com via Yahoo.