DeepSeek V4.1-Flash-1017: Open Weights, Cheaper Agents, and What the Benchmarks Actually Say

Published: October 18, 2026 · 8 min read

On October 17, 2026, DeepSeek quietly updated its fastest, cheapest model. The new checkpoint — DeepSeek-V4.1-Flash-1017 — arrives with its weights published under the MIT license, a 1-million-token context window, and DeepSeek's claim of improvements across all 11 benchmarks it reports. The headline that matters most isn't a score, though. It's the combination: an open-weight model you can download and modify, priced low enough that running it as a coding agent costs cents per task instead of dollars. That pairing — genuinely open, genuinely cheap — is rare, and it's why this release deserves a closer look than the average model update.

What actually shipped

The release note and model card describe V4.1-Flash-1017 as "a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks." Both the modesty of the label and the confidence of the numbers are worth noting — and taking on DeepSeek's terms. The update went live the same day on DeepSeek's website, its app, and its API, with the weights posted to Hugging Face and ModelScope.

The technical profile is unchanged in shape from V4.1-Flash: a 1-million-token context window, image input alongside text, tool calling, and adjustable reasoning effort. The Hugging Face model card lists 763 billion parameters with weights stored in a mix of FP8 and BF16 precision, and DeepSeek recommends a context of 1M tokens, a maximum output of at least 256K tokens, and temperature and top_p both set to 1.0. On the API, the deepseek-flash identifier now serves the 1017 checkpoint — existing integrations get the new model with no code changes. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work and are routed to the new checkpoint at the same Flash price. One operational footnote: the update shipped under the existing API identifier with no changelog at first, so existing API consumers absorbed the behavior change before documentation appeared. If you're pinned to deepseek-flash in production, that silent swap is how it went live.

The numbers — and why to hold them loosely

Here are DeepSeek's reported figures, with the essential caveat up front: every score below comes from DeepSeek's own tests on harnesses it chose, and none has been independently reproduced as of publication. Treat these as claims, not facts.

On DeepSWE v1.1, a software-engineering benchmark, the score rises from 74.2 to 77.0. Terminal-Bench 4.0 jumps from 31.2 to 42.3, and ProgramBench from 20.3 to 29.1. MarkTechPost, working from DeepSeek's published tables, reports the new checkpoint beats the previous Flash on all 11 benchmarks by 2.0 to 11.1 points. In DeepSeek's own comparative tables, the new model beats Kimi K3 on 10 of 11 tests and GPT-5.6 Sol on nine, but trails Claude Opus 5 on seven — on raw reasoning, Opus still leads comfortably (Humanity's Last Exam: 42.6 for Flash-1017 versus 56.3 for Opus).

What's interesting about the pattern, not just the heights: the biggest gains are on agentic coding and automation benchmarks — Terminal-Bench, ProgramBench, Apex Bench — with smaller moves on static reasoning. That's consistent with what DeepSeek says it trained for: reinforcement learning with self-feedback to build complete projects, shown in demos like a web racing game and 3D scenes. If that pattern holds under independent testing, it means the model is specifically getting better at the multi-step work agents actually do, rather than just at answering questions.

The real story: cost per finished task

Price per token barely registers in most people's heads. Cost per finished task does — and that's the number DeepSeek is pushing hardest. On DeepSeek's own serving, pricing is $0.15 per million input tokens off-peak ($0.30 at peak) and $0.60 per million output tokens off-peak ($1.20 at peak), with cache reads at a near-zero $0.003 per million. Against that price, DeepSeek's chart plots DeepSWE scores against task cost: the new model at its top setting costs just under $0.20 per task, versus about $0.50 for the previous version — and rival models in the same chart cost between $2 and more than $10 per task.

Two levers drive that number down. First, the price itself is simply low — input costs roughly a quarter of what DeepSeek's own V4 Pro charges, and output about a third. Second, and more unusual, DeepSeek says the new checkpoint reduces task execution time and output length: shorter answers mean fewer billed tokens per finished job, and faster completion means less latency per task. A lower-effort setting reportedly still scores above 72% at roughly $0.10 per task. If you run coding agents in bulk — repo migrations, test sweeps, documentation passes — the difference between $0.10–$0.20 and $2–$10 per task is the difference between a pipeline you run routinely and one you budget carefully. Again: these figures are DeepSeek's. The pricing is independently checkable on DeepSeek's site; the benchmark-cost claims are not yet.

Open weights under MIT: why that changes the calculus

The weights release is the part with the longest half-life. The repository and the weights are released under the MIT License, which permits commercial use and modification — no special commercial terms, no seat limits, no per-user API toll on top of the license. For a team pricing a switch away from OpenAI or Anthropic, MIT licensing removes the open-versus-closed tradeoff that usually forces the decision: you can self-host, fine-tune, distill, or just inspect the model without negotiating anything.

The practical economics of self-hosting deserve a clear-eyed look, though. The weights reportedly total about 510 GB — that's the download, and then you need the GPU memory to actually serve them, plus engineering time to run inference well. Self-hosting a 763B-parameter model is not a weekend project; it's a real infrastructure commitment. The open weights matter most for organizations that already run inference stacks, researchers who want to study or distill the model, and anyone with compliance reasons to keep data off third-party APIs. For everyone else, the API at these prices is almost certainly cheaper than the GPU bill. The point of the MIT release isn't that everyone will self-host — it's that everyone can, and that option disciplines the API pricing.

Two new APIs arrive alongside the model

DeepSeek also opened two tools to all API users: a Search API that returns ranked web results for a query, and a Fetch API that returns a single page as Markdown — both built on DeepSeek's in-house search engine, which the company says serves hundreds of millions of requests daily. Each costs $0.75 per 1,000 successful requests (failures and empty results are free), and they're capped at 5 requests per second per account. Calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex stay free until November 1, 2026.

This is DeepSeek's answer to the browsing-tool problem: rather than each agent implementing its own janky search wrapper, the provider sells a first-party search endpoint with operators like site:, lang:, after:, and before:, plus rerank scores and authority tiers on each result. At $0.75 per thousand calls, it's priced as infrastructure, not as a premium feature. For agent builders, it reduces one integration to an API call. For everyone else, it's optional — worth knowing about, not worth restructuring around until you've tested whether the results beat whatever you currently use.

Who should actually care

Agent builders and automation teams: this is the most natural audience. If you're running coding agents or automation pipelines where per-task cost is the constraint, a $0.10–$0.20 per-task option with a 1M context window and tool calling deserves a real evaluation — benchmark it yourself on your own tasks before migrating, since DeepSeek's numbers are DeepSeek's.

Teams wanting open weights without the usual tradeoffs: MIT licensing on a current-generation model removes the license-compliance headache that comes with custom "open-ish" licenses. If your policy requires genuinely permissive weights, this is now one of the strongest options available.

Individual developers and hobbyists: via the API, this is one of the cheapest ways to run a capable coding model with a huge context window. A million-token prompt costs $0.15 off-peak — cheap enough to experiment freely. You won't be self-hosting 510 GB of weights, and that's fine; the API is the path.

Everyone else: probably not much changes. If you use AI through chat apps for writing, questions, and everyday tasks, a new developer-oriented checkpoint won't be visible to you. The models you'll actually meet got an incremental upgrade, not a new personality.

What to watch out for

Vendor benchmarks are marketing until reproduced. DeepSeek's numbers look strong — but they're run by DeepSeek, on harnesses DeepSeek chose, comparing against scores DeepSeek computed for rivals. Independent reproductions take weeks. The honest position is: promising, unverified.

The silent API swap. Serving a new checkpoint under an existing API identifier with no initial changelog is efficient and also a little alarming. If you build production systems on deepseek-flash, consider whether you're comfortable with behavior changing underneath you without notice — or pin versions where the provider allows it.

KV cache compression is the technical bet — watch it. The companion paper, titled "DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression," puts the memory story on the cover: roughly 890 bytes per token, meaning a full 1M-token context needs under a gigabyte of cache in GPU memory. That's a genuine engineering advance if it holds — and it arrives in the middle of a run of inference-side KV cache research over the past month. But aggressive compression can have subtle quality costs at the extremes of context length; long-context claims are exactly the kind that need independent verification at the claimed 1M tokens, not just at benchmark lengths.

It's still a Chinese-developed model with a Chinese-developed stack. DeepSeek publishes weights and code openly, which mitigates a lot — but if your threat model includes supply-chain concerns about model weights from any particular country, the license doesn't change the origin. Evaluate it the way you'd evaluate any open-weight release: download, test, decide.