DeepSeek’s Flash Model Achieves Net Gains in Cost and Agentic Performance
DeepSeek has released V4.1-Flash-1017, an open update on Hugging Face and ModelScope. The model promises better scores in coding and agentic tasks, with shorter and less costly outputs.
DeepSeek launched DeepSeek-V4.1-Flash-1017 on October 17, 2026, on its website, app, API, and as open weights on Hugging Face and ModelScope. This Flash model version improves coding and agent tasks while reducing cost per task, a key point for developers and companies automating entire workflows.
A Faster, Shorter, More Cost-Effective Flash
DeepSeek’s message is simple: the update advances on multiple fronts simultaneously. According to the technical report, V4.1-Flash-1017 reduces execution time and response length, lowering compute costs and improving the performance-price tradeoff, which the company describes as a "Pareto improvement." In other words, the model does more useful things without requiring more resources. Not bad for a “Flash” that precisely avoids pyrotechnic effects.
The company says it trained this model with reinforcement from self-feedback, a method called RLSF, for reinforcement learning with self-feedback. The principle: the model rewards itself based on its own confidence scores or internally generated critiques, without human evaluators or external reward models. DeepSeek claims this approach helps in building complete projects, from video games to 3D modeling and design.
Visible Gains in Coding and Agents
In its own tests, DeepSeek-V4.1-Flash-1017 outperforms the previous version across all evaluated categories. On DeepSWE v1.1, it reaches 77.0% versus 74.2% for DeepSeek-V4.1-Flash, with an average cost below 0.20 dollars per task. In low-cost mode, it remains above 72% for about 0.10 dollars. The former model ran around 0.50 dollars per execution.
The technical document also places the model ahead of several larger rivals in certain categories. DeepSeek notably cites Claude Opus 5, GPT-5.6-Sol, Kimi K3, and GLM 5.3 on coding and agent benchmarks. On Terminal-Bench 4.0, the model scores 42.3, versus 31.2 for its predecessor. On NL2Repo-Bench, it climbs to 68.2, up from 65.4 previously. These figures remain internal for now, and independent benchmarks have yet to be published [to be verified].
Open Search APIs Available to All Developers
The news doesn’t stop at the model. DeepSeek also opens its Search & Fetch APIs to all developers, allowing them to query its internal search engine. The company says this engine already handles hundreds of millions of queries daily worldwide. Each endpoint call is priced at 0.75 dollars per 1,000 calls.
For some uses, access remains free until October 31, 2026, 16:00 UTC via server-side tools in DeepSeek Harness, Claude Code, and Codex. DeepSeek also specifies that the model is optimized for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. Regarding distribution, the repository is published under the MIT license, with inference recommendations mentioning a context window of 1 million tokens and a maximum output of at least 256,000 tokens.
What This Update Really Changes
This version confirms DeepSeek’s strategy: pushing a more compact model but aggressive on cost-performance ratio, and connecting it directly to development tools. For teams building software agents, the gain is not just academic. Lower costs, shorter outputs, and broader integration into work environments can change the final bill, especially at scale.
One unknown remains. The document presents solid internal results, but the real gap with competing models will depend on independent tests and real-world use. Between a good benchmark curve and smooth product integration, there is often a small world — one that developers fill with bugs, delays, and cold coffee.
Key points
- DeepSeek-V4.1-Flash-1017 released on October 17, 2026.
- The model is available on site, app, API, and Hugging Face.
- DeepSWE v1.1: 77.0% at about 0.20 dollars.
- Search & Fetch APIs cost 0.75 dollars per 1,000 calls.
- DeepSeek cites Claude Opus 5 and GPT-5.6-Sol.
- The internal engine handles hundreds of millions of queries daily.
By the numbers
- 77.0% — DeepSWE v1.1 score, DeepSeek, October 2026.
- 0.20 dollars — average cost per agentic task, DeepSeek, October 2026.
- 0.75 dollars — price for 1,000 Search & Fetch API calls.
- 1 million tokens — recommended context window, DeepSeek technical report.
- October 31, 2026 — end of announced free access for certain tools.