Chinese AI company DeepSeek has released DeepSeek-V4.1-Flash-1017, an updated version of its Flash model. The new model brings improvements in coding, AI agent tasks, speed, and efficiency.
The model is available through DeepSeek’s website, mobile app, and API. Its model weights have also been published on Hugging Face and ModelScope.
DeepSeek says the updated model performs better on benchmarks than the previous version while completing tasks faster and producing shorter responses. This helps reduce costs without sacrificing performance.
New Training Method Improves Coding
DeepSeek has introduced a post-training method called reinforcement learning with self-feedback (RLSF) to improve the model’s software development abilities.
This method allows the AI to use its own confidence scores or generated feedback as learning signals instead of depending entirely on human evaluators or external reward models.
According to DeepSeek, this approach helps the model build complete projects from scratch, including games, 3D models, and design projects. It can also create short animated explainer videos.
The company has optimized V4.1-Flash-1017 for development tools such as DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy. DeepSeek recommends using version v0.2.1-rc.1 or later of DeepSeek Harness.
Smaller Model Delivers Strong Results
According to DeepSeek’s internal benchmarks, V4.1-Flash-1017 performs better than its predecessor across all tested categories.
The company also claims that the smaller model outperforms several larger AI systems, including Claude Opus 5, GPT-5.6 Sol, Kimi K3, and GLM 5.3, on selected coding and AI agent tests.
On the DeepSWE benchmark, the model achieved a score of 77.0% at an average cost of less than $0.20 per task.
At a lower performance setting, it scored above 72% at a cost of around $0.10 per task. The previous version scored 74.2% at approximately $0.50 per run.
DeepSeek’s comparison also shows some competing models costing between roughly $2 and more than $10 per task, with lower scores on the selected benchmark.
However, these results come from DeepSeek’s own testing. Independent evaluations and real-world developer experiences may produce different results.
DeepSeek Opens Its Search Tools to Developers
Alongside the model release, DeepSeek is making its internal web search technology available to outside developers for the first time.
The company has introduced two tools: Search API and Fetch API. These services use DeepSeek’s search engine, which reportedly processes hundreds of millions of queries worldwide each day.
Both APIs are priced at $0.75 per 1,000 calls.
DeepSeek API users can also access the tools for free through server-side integrations in DeepSeek Harness, Claude Code, and Codex until October 31, 2026, at 4:00 PM UTC.
China May Introduce New Rules for Open-Weight AI Models
The release comes as the Chinese government reportedly considers stricter rules for publishing open-weight AI models and other important AI technologies.
Under the reported proposal, a model’s position among the 10 most capable AI systems could be one factor in deciding whether additional restrictions apply.
If these rules are introduced, they could affect the future release and availability of Chinese open-weight AI models and make it more difficult for developers to rely on them.