DeepSeek V4.1 Flash checkpoint reports 11-point Terminal-Bench jump under MIT license

What shipped

DeepSeek has released DeepSeek-V4.1-Flash-1017, a new checkpoint that the model card itself describes as a minor update. According to a post on dev.to that walks through the release, it is a 763-billion-parameter model distributed in BF16, F32, F8 and I8 weights, with a one-million-token context window and a recommended maximum output of at least 256K tokens. The weights are MIT licensed with commercial use permitted, the model runs on Transformers, vLLM and SGLang, and community builds already include 33 quantized variants for llama.cpp, Ollama and LM Studio. The stated goal of the update is faster task execution and shorter outputs, improving the trade-off between performance and cost.

The reported benchmark gains

The dev.to post reproduces the vendor's comparison between the new checkpoint and the previous V4.1 Flash. Terminal-Bench 4.0, an agentic benchmark for terminal work, rises from 31.2 to 42.3 — an 11.1-point jump. Humanity's Last Exam without tools improves from 36.8 to 42.6, DeepSWE v1.1 from 74.2 to 77.0, and Automation-Bench from 54.8 to 60.5.

The post also notes the reception: a Hacker News thread reportedly reached 849 points and more than 600 comments within hours, with much of the discussion questioning whether the Terminal-Bench figure could be genuine.

Caveats on the numbers

These are DeepSeek's own reported figures, and the post is explicit about that limitation. It points out that independent evaluators are still catching up with the new checkpoint, and that earlier third-party runs of V4.1 Flash by Artificial Analysis and Vals.ai placed the previous version at the top of open-weight models while showing it trailing on specialised evaluations, including SRE-style and legal-agent tasks. The post's recommendation is to wait for independent reruns before making architectural decisions based on a model card.

A second release: search and fetch APIs

Buried in the same news cycle, according to the post, DeepSeek launched search and fetch APIs priced at $0.75 per 1,000 successful requests, billed only when a call returns HTTP 200 with actual results. A POST to the search endpoint returns ranked results, summaries and optionally page content, while the fetch endpoint returns a page converted to Markdown. The APIs support site, lang and date-range operators and are capped at five requests per second per API, account-wide. Commenters cited in the post compare that pricing with roughly $4 per 1,000 requests at Exa, $5 at Brave and $12 at Kagi, though whether pages behind anti-bot protections return usable content remains an open question.

Running it yourself

At 763 billion parameters, this is not a laptop-scale model. The post notes that the quantized builds make local hosting possible but not inexpensive, and that for most teams the API is the realistic path. Listed pricing puts V4.1 Flash at roughly $0.30 per million input tokens and $1.20 per million output tokens, with off-peak discounts on top, though the post cautions that rates change and should be checked against the provider's page.

Why it matters

The largest gain in this release is on agentic terminal work — the repetitive loops of reading files, editing code, running commands and retrying after failures that consume the most tokens in production. If shorter outputs and fewer steps to completion hold up under independent testing, the cost advantage compounds even at unchanged per-token prices. Combined with an MIT license that removes lock-in risk, an open-weight model delivering a measurable agentic improvement in what it calls a minor update applies real pricing pressure on closed-model vendors, whose differentiation increasingly rests on distribution and tooling rather than raw capability. The sensible posture, as the post suggests, is to run the checkpoint against your own agent traces — measuring steps-to-completion and total tokens, not just accuracy — while holding out for third-party benchmark confirmation.