Skip to content

DeepSeek V4.1-Flash-1017: Open Weights, Faster Agents, Same Low Price (October 2026)

    DeepSeek doesn’t do slow weeks anymore. Yesterday — Saturday, October 17 — the company quietly shipped DeepSeek-V4.1-Flash-1017, an updated checkpoint of its V4.1 Flash model, and posted the full weights on Hugging Face under the MIT license the same day. No new pricing, no new API name, no new architecture. And yet, in its own benchmark tables, this “minor update” posts some of the biggest agentic-coding jumps of the month.

    There’s a catch you’ll want to keep in mind as you read: every number below comes from DeepSeek’s own release material. Nobody has independently reproduced them yet. That’s not a reason to dismiss them — it’s a reason to read them like a lab notebook written by the lab.

    What Actually Shipped

    Here’s the clean version, straight from DeepSeek’s release note and model card, all dated October 17, 2026:

    • An updated checkpoint, not a new model. DeepSeek calls 1017 “a minor update” to V4.1 Flash, which launched on September 10. The API model string doesn’t change — if you already call deepseek-flash, you’re already on 1017. The calling method is unchanged.
    • Open weights, day one. The checkpoint is published on Hugging Face (and ModelScope) under the MIT license, which permits commercial use. The repo went up at 02:40 UTC on October 17. You can run the hosted model and your own deployment side by side and compare them honestly — that’s rare, and genuinely useful.
    • Same price. $0.15 per million input tokens, $0.60 per million output, per the OpenRouter listing. The legacy names (deepseek-v4-flash, deepseek-v4-flash-vision-exp) still work and are billed at the Flash price.
    • 1-million-token context (1,048,576 tokens), text and image in, text out.
    • Tuned for specific harnesses. DeepSeek says the checkpoint is optimized for its own DeepSeek Harness (v0.2.1-rc.1 or later), Tencent’s CodeBuddy, and WorkBuddy — a reminder that “benchmark numbers” are often “benchmark numbers in the vendor’s favorite harness.”

    The companion technical report carries an interesting title: “DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression.” Translation: a big part of this update’s story is memory efficiency. One outlet worked through the numbers and found that at DeepSeek’s stated 890 bytes per token, a full 1M-token prompt needs under a gigabyte of cache in GPU memory. The weights themselves are the heavier bill — around 510 GB. If you were hoping to self-host this on a single consumer GPU, that number is your reality check.

    The Benchmarks: Impressive, With Asterisks

    DeepSeek’s table shows 1017 beating the previous Flash on all 11 benchmarks it reports, with gains from 2.0 to 11.1 points:

    • DeepSWE v1.1 (coding agents): 77.0, up from 74.2 — ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0) on this specific test.
    • Terminal-Bench 4.0: 42.3, up from 31.2 — a 36% jump, though still behind Opus 5’s 51.8 in the same table.
    • ProgramBench: 29.1 (up 8.8 points), Apex Bench v1.1: 72.1, HLE without tools: 42.6 (up from 36.8).

    Now the asterisks, because truth-first means saying them out loud:

    1. These are DeepSeek’s numbers, produced on test harnesses it chose. No independent run has been published yet. Treat them as a strong signal, not a settled fact.
    2. “On par with Opus 5” is a mood, not a measurement. Some early chatter claimed 1017 was roughly on par with Claude Opus 5. DeepSeek’s own table disagrees: Opus 5 scores 56.3 vs 42.6 on HLE without tools and 51.8 vs 42.3 on Terminal-Bench 4.0. The 1017 leads on specific agent tests; it does not dethrone Opus 5 across the board. Show me the eval, not the vibe.
    3. The honest headline is cost-per-task. One independent pilot (56 trials) found 1017 completing tasks at about $0.0157 versus $0.0577 for GPT-6.1 Sol and $0.0653 for Claude Sonnet 5.5. DeepSeek itself advertises “under $0.20 per task” — that’s company marketing, but the direction is real: shorter outputs + lower prices = dramatically cheaper agent runs.

    The Bigger Move: DeepSeek’s Own Search API

    The part of this release that got less attention may matter more long-term. DeepSeek opened its Search and Fetch APIs to all developers — a web search layer powered by its in-house search engine, which it says serves hundreds of millions of requests a day.

    • $0.75 per 1,000 successful calls, per endpoint.
    • Free until billing starts on November 1, 2026 for calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex.
    • Tuned for DeepSeek’s own Harness, Tencent CodeBuddy and WorkBuddy.

    This is DeepSeek moving up the stack: a cheap, open model plus a first-party search layer behind one API key. Agent builders no longer need to wire in a separate search provider for retrieval. That puts DeepSeek in direct competition with dedicated search-API vendors — and it’s the clearest sign yet that the company is building toward a full agent-infrastructure play, not just selling tokens.

    Who Should Care?

    • Agent builders already on deepseek-flash: you’re on 1017 automatically — re-run your own coding/agent tests before trusting September results against October reality.
    • Self-hosters: MIT-licensed weights, 510 GB download, sub-gigabyte cache for 1M-token prompts. Compare it against the hosted version honestly; the weights are there precisely so you can.
    • Everyone else: the October 2026 pattern continues. First Liquid AI’s tiny LFM2.5 MoE, then Meta’s 816B Muse Spark Open, now a leaner DeepSeek checkpoint. The frontier stopped being a single race to the biggest parameter count — it’s fragmenting into specialists and efficiency plays. 1017 is one of the cheapest per-task agents on the market right now, and that matters more to your bill than any parameter count ever will.

    Just remember: the scores are DeepSeek’s. The price is real, the weights are open, and the search API is genuinely new. Everything else is a claim waiting for an independent run.