TL;DR: DeepSeek V4.1-Flash-1017 is a “minor update” to V4.1-Flash that the company says scores higher on coding and agent benchmarks while finishing tasks faster with shorter outputs. The weights are on Hugging Face under the MIT License, the API name deepseek-flash now points to it at unchanged prices, and DeepSeek has opened its in-house web search to developers through new Search and Fetch APIs.
DeepSeek V4.1-Flash-1017 landed on October 17, 2026, and it is the kind of release that matters more to builders than to chatbot fans. According to DeepSeek’s API change log, the update replaces V4.1-Flash on the DeepSeek website, app and API, with the calling method and pricing unchanged. The open weights went up the same day on Hugging Face, and the company paired the model with something it has never offered outside its own products: public access to its search engine.
Here is what changed, what the numbers do and don’t tell you, and who should bother switching.
What DeepSeek actually shipped
The Hugging Face model card describes V4.1-Flash-1017 as “a minor update to DeepSeek-V4.1-Flash, delivering substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks.” DeepSeek also says the model “reduces task execution time and output length,” which is the part that should get developers’ attention: shorter outputs mean fewer billed tokens per finished job.
- License: the repository and the weights are released under the MIT License, which allows commercial use and modification.
- Size: the Hugging Face repo lists the checkpoint at 763B parameters and tags it as an image-text-to-text model, consistent with V4.1-Flash’s native visual understanding.
- Context: DeepSeek’s recommended settings are a 1M-token context window,
max_tokensof at least 256K, and temperature and top_p both at 1.0. - API: per DeepSeek’s quick-start docs,
deepseek-flashis now served by V4.1-Flash-1017. The retired namesdeepseek-v4-flashanddeepseek-v4-flash-vision-expstill work and are routed to the new model at the Flash price. - Tooling: DeepSeek says it specifically tuned the model for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, and recommends Harness 0.2.1-rc.1 or later.
The benchmark numbers (DeepSeek’s own)
Every figure below comes from DeepSeek’s model card and change log. Independent results were not available at the time of writing, so treat them as the company’s claims rather than settled rankings.
| Benchmark | V4.1-Flash-1017 | V4.1-Flash (Sept) |
|---|---|---|
| HLE (no tools) | 42.6 | 36.8 |
| Terminal-Bench 4.0 | 42.3 | 31.2 |
| DeepSWE v1.1 | 77.0 | 74.2 |
| ProgramBench | 29.1 | 20.3 |
| NL2Repo-Bench | 68.2 | 65.4 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 |
| Agents’ Last Exam | 35.4 | 31.8 |
| Apex Bench v1.1 | 72.1 | 64.6 |
| BabyVision (with tools) | 92.5 | 89.6 |
The model card also lists rival scores. On DeepSeek’s chart, the new Flash beats Claude Opus 5 on DeepSWE (77.0 vs 74.0) and Automation-Bench (60.5 vs 50.3), but trails it on HLE (42.6 vs 56.3), Terminal-Bench 4.0 (42.3 vs 51.8) and ProgramBench (29.1 vs 37.0). DeepSeek notes it ran Terminal-Bench 4.0 in the Claude Code harness and DeepSWE in the Mini-SWE-V2 harness, so scores from other setups may not line up.
Why the cost story is the real headline
The Decoder’s write-up of DeepSeek’s cost chart puts numbers on the “shorter outputs” claim: on DeepSWE, V4.1-Flash-1017 reaches 77.0% at an average of just under $0.20 per task, and about 72% for roughly $0.10 on its low setting, while the previous Flash scored 74.2% at around $0.50. The same chart places several closed rivals between $2 and more than $10 per task.
If those ratios hold up in independent tests, the upgrade is less about a few benchmark points and more about running agent loops (code fixes, repo-wide refactors, automation scripts) at a fraction of what frontier APIs charge. That’s the same pressure we flagged when Reflection opened Beam 501B and when Meta released Muse Spark Open: open-weight labs are now competing on cost per finished task, not just raw scores.
DeepSeek’s search engine is now an API
The second announcement may outlast the model. DeepSeek’s change log says its web and API search runs on an in-house engine that “reliably handles hundreds of millions of search requests from around the world every day,” and that it is now open to developers through Search and Fetch endpoints. Search returns ranked web results; Fetch pulls a single page back as Markdown, which is what agents and retrieval pipelines need.
What it costs
- $0.75 per 1,000 calls for each endpoint, according to The Decoder’s report on DeepSeek’s announcement.
- Free for now through server-side tools in DeepSeek Harness, Claude Code and Codex for DeepSeek API users, until October 31, 2026 at 16:00 UTC, per the same report.
- Both use the existing DeepSeek API key and base URL, so adding live web results to an agent is mostly a configuration change.
Running it yourself: what to expect
Open weights don’t mean easy hosting. DeepSeek’s inference README calls its code “a readable reference implementation rather than a production serving engine.” The example conversion script splits the weights into one file per tensor-parallel rank (the sample uses eight) with experts stored in FP4, which points to multi-GPU server hardware, not a laptop.
There is also no Jinja chat template. Instead, DeepSeek ships a Python reference encoder plus deepseek-recipe, a set of Rust libraries with Python bindings that convert Chat Completions, Messages and Responses-style requests into DeepSeek’s prompt format and parse outputs back, including tool calls and images. Inference, tool execution and HTTP transport are left to you.
The policy cloud over open Chinese models
The release lands a day after China published draft export controls covering advanced AI models, agents and chip designs, which we cover in China’s new AI export rules. Reports on the draft say open-source releases are exempt, but companies building long-term products on Chinese open-weight models should watch how the final rules treat top-ranked systems. DeepSeek’s own funding plans, covered in our report on its reported $12 billion round, make it one of the companies with the most at stake.
Who should switch
- Already on the DeepSeek API: you’ve been switched automatically if you call
deepseek-flashor the legacy Flash names. Re-run your evals, since shorter outputs can change downstream parsing. - Building coding or automation agents: worth a head-to-head test on your own tasks, given DeepSeek’s cost-per-task claims.
- Self-hosters: budget for serious GPU capacity and your own serving stack, or wait for third-party inference providers.
The bottom line
DeepSeek V4.1-Flash-1017 is a point release with outsized implications: MIT-licensed weights, company-reported gains on coding and agent benchmarks, noticeably cheaper tasks on DeepSeek’s own charts, and a new Search and Fetch API that turns its search engine into agent infrastructure. Wait for independent benchmarks before you rip out another model, but if cost per completed task is your bottleneck, this is the release to test this week.
The Teck Hustlers Editorial Team covers technology, business, finance and gaming news for readers who want the story behind the headline. Every article is researched, written and edited by our in-house team, checked against company announcements, filings and reputable reporting, and updated when new facts emerge. We focus on what each development means for founders, professionals, investors and everyday users. Have a tip or a correction? Reach us through our contact page.
