DeepSeek V4.1 Flash 1017: Bigger Agent Gains, Same Price

DeepSeek has shipped a mid-cycle upgrade to its smallest frontier model. On October 17 it released DeepSeek-V4.1-Flash-1017, a refreshed checkpoint of the V4.1 Flash model that launched in September. It brings large gains on coding and agent benchmarks, finishes tasks faster, and keeps the same price. DeepSeek also opened its in-house search engine to developers as a paid API. The new weights are already on Hugging Face under the MIT license, so this is an open-weight frontier upgrade that anyone can download, study and build on.

• Release date: October 17, 2026, live in the DeepSeek web app, mobile app and API under the existing deepseek-flash model name
• Benchmark jumps (DeepSeek's figures): Terminal-Bench 4.0 rises from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and Apex Bench v1.1 from 64.6 to 72.1
• Open weights: the repository and weights are MIT-licensed, with a 1M-token context window and image plus text input
• New Search API: DeepSeek's own search engine is now available to developers, priced at $0.75 per 1,000 successful requests according to Trending Topics

What Changed in DeepSeek V4.1-Flash-1017?

DeepSeek describes 1017 as a "minor update" to V4.1 Flash, but the benchmark table on its model card is anything but minor. Compared with the September checkpoint, DeepSeek reports:

• Terminal-Bench 4.0: 42.3, up from 31.2 (run in the Claude Code harness)
• DeepSWE v1.1: 77.0, up from 74.2
• Humanity's Last Exam (no tools): 42.6, up from 36.8
• Automation-Bench v1.0.6: 60.5, up from 54.8
• Agents' Last Exam: 35.4, up from 31.8
• BabyVision (with tools): 92.5, up from 89.6

The gains cluster around agentic coding and long, multi-step tasks, which is where developers increasingly use open models. All of these figures come from DeepSeek itself. As of October 18, the independent Artificial Analysis leaderboard had not yet tested the new checkpoint, so treat any head-to-head rankings as provisional until third parties weigh in.

Faster, Shorter, Cheaper Agent Runs

The quieter improvement may matter more in practice. DeepSeek says 1017 cuts both task runtime and output length. It calls the result a new "performance-cost Pareto optimum" across several agent harnesses. In plain terms, the model reaches better answers while writing fewer tokens to get there. Since API billing is per token, shorter outputs mean lower bills even though the price per token has not changed.

DeepSeek tuned the checkpoint specifically for its own DeepSeek Harness, plus Tencent's CodeBuddy and WorkBuddy. It recommends upgrading to DeepSeek Harness 0.2.1-rc.1 or later. Trending Topics reports API pricing holding at $0.30 per million input tokens and $1.20 per million output tokens.

Is DeepSeek V4.1-Flash-1017 Open Source?

Yes. The Hugging Face repository states that both the code and the model weights are licensed under MIT, one of the most permissive licenses available. The model card recommends a 1M-token context window, max output of at least 256K tokens, and temperature and top_p both set to 1.0. It also ships a reference prompt encoder and a Rust-based toolkit, deepseek-recipe, that converts standard Chat Completions and Responses API requests into DeepSeek's prompt format.

If you followed our coverage of the original DeepSeek V4.1 Flash open-weights release, the upgrade path is simple. API users calling deepseek-flash were moved to 1017 automatically, and self-hosters can pull the new checkpoint.

DeepSeek Search API Opens for Developers

The second launch gives developers DeepSeek's in-house search engine, the same one behind its consumer app. DeepSeek says it indexes hundreds of billions of web pages and handles hundreds of millions of searches a day. It also says the engine filters out SEO spam and low-quality AI-generated content. Developers can use it to fetch search results or to retrieve individual web pages as text, a core building block for research agents. That puts DeepSeek alongside existing agent search providers.

Why This Open-Weight Upgrade Matters

Open-weight models keep getting better between major releases, and 1017 is a clear example. Developers get a meaningfully stronger coding and agent model at the same price, with weights they can run on their own hardware. Explore more open-model news in our AI coverage and see how it compares with Reflection Beam's Apache-licensed open weights.

Sources: DeepSeek API Docs Change Log — October 17, 2026; DeepSeek-V4.1-Flash-1017 model card, Hugging Face — October 17, 2026; Trending Topics — October 18, 2026.