DeepSeek Releases V4.1-Flash-1017 With MIT-Licensed Weights and New Search and Fetch APIs

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, 2026, a small update to its low-cost Flash model that the company says improves reasoning, coding, agent, and vision benchmark scores while finishing tasks faster with shorter outputs. The weights are published on Hugging Face and ModelScope under the MIT license, which permits commercial use, modification, and redistribution. On the same day, DeepSeek opened two new developer tools, a Search API and a Fetch API, to all of its API users.

The Flash line is DeepSeek’s budget tier: models meant for jobs you run thousands of times a day, such as coding agents, summarization, and classification. The previous version, V4.1-Flash, launched on September 10, 2026. DeepSeek describes the new checkpoint as a minor update, but its published benchmark table shows gains on all 11 tests it reports, from about 2 points on ZeroBench to 11.1 points on Terminal-Bench 4.0. On DeepSWE v1.1, a software-engineering benchmark, the score rose from 74.2 to 77.0, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 73.0 in the same comparison. On Terminal-Bench 4.0 the score climbed from 31.2 to 42.3, and on ProgramBench from 20.3 to 29.1. In the table, 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, including Humanity’s Last Exam, where Opus 5 scores 56.3 against the Flash update’s 42.6. All of these figures come from DeepSeek’s own tests; no independent benchmark reproductions were available at publication time.

Cost is the headline claim. DeepSeek calls the release a Pareto improvement in performance and cost, and its chart of DeepSWE scores against cost per task puts the new model at just under $0.20 per task on its top setting, down from about $0.50 for the previous version. A lower-effort setting still scores above 72% at roughly $0.10 per task. The same chart places several rival models between $2 and more than $10 per task. API pricing is unchanged: for prompts up to 100,000 tokens, $0.15 per million input tokens off-peak and $0.30 at peak, with $0.60 and $1.20 per million output tokens. Cached input costs $0.003 off-peak and $0.006 at peak per million tokens. The model went live under the existing API name deepseek-flash on DeepSeek’s website, app, and API, so existing API users are switched automatically with no code changes. The older aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired and now route to the new model at Flash prices.

The model card describes a mixture-of-experts design, and the context window is 1 million tokens with up to 384,000 tokens of output. The model offers thinking and non-thinking modes, tool calls, and image input, and DeepSeek says it tuned the update for its own DeepSeek Harness agent framework as well as Tencent’s CodeBuddy and WorkBuddy, recommending DeepSeek Harness v0.2.1-rc.1 or later. DeepSeek says it trained the update with reinforcement learning from feedback on its own attempts, and its example outputs include a 3D model of the Forbidden City and a Portal 2 recreation built in the Godot engine.

Beyond the model itself, the bigger move may be infrastructure. DeepSeek opened its Search API, which finds web pages with the company’s own in-house search engine, and its Fetch API, which returns a page as Markdown, to every API user. Each costs $0.75 per 1,000 successful calls. Calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex stay free for DeepSeek API users until October 31, 2026. With a low-cost agent model and a first-party web search layer behind one API key, DeepSeek is positioning itself against dedicated search API providers for agent workloads. The release lands as DeepSeek’s latest funding round has reportedly grown to between $12 billion and $15 billion at a valuation of about $75 billion, ahead of a planned Shanghai IPO, according to CNBC.

Why it matters: open-weight labs are now competing on cost per finished task rather than raw benchmark scores alone. DeepSeek’s MIT-licensed weights remove the open-versus-closed tradeoff for teams pricing a switch away from closed providers, and the per-task cost claims, if independent tests confirm them, would make Flash-class models the cheapest way to run coding agent loops at scale. Readers should treat the benchmark gains as the company’s claims until third-party evaluations reproduce them.