Deepseek has released Deepseek-V4.1-Flash-1017, an updated model aimed at software development and AI agent tasks. The company says it improves benchmark performance while running faster and producing shorter outputs, a combination that could lower the cost of completing a task.
A model built for end-to-end development
Deepseek says the model is designed to handle complete software projects, including work that starts from scratch. The company describes using reinforcement learning with self-feedback, or RLSF, during post-training. In this approach, the model can use its own confidence scores or critiques it generates as a reward signal.
According to Deepseek, this removes the need for human evaluators or external reward models in that training process. The company says the approach supports tasks such as game programming, 3D modeling, and design work. It also says the model can animate compact explainer videos.
The model is tuned for development environments including Deepseek Harness, Tencent CodeBuddy, and WorkBuddy. Deepseek recommends version v0.2.1-rc.1 or later for Deepseek Harness.
Benchmark results come with a caveat
In benchmarks run by Deepseek, V4.1-Flash-1017 scores better than its predecessor, V4.1-Flash. The company also reports that the new model outperforms larger competitors, including Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3, in several coding and AI agent categories.
One DeepSWE benchmark comparison focuses on the relationship between task score and cost. Deepseek reports a score of 77.0 percent for V4.1-Flash-1017 at an average cost of just under $0.20 per task. In a low setting, it scores above 72 percent at about $0.10 per task.
For comparison, the predecessor scored 74.2 percent at around $0.50 per run. Deepseek's chart places GPT-6 Astra, Gemini 3.8 Flash, and Claude Opus 5 between $2 and over $10 per task, with none matching the smaller model's reported accuracy on that chart.
These figures are the company's own results. Independent benchmarks have not been released, and benchmark scores do not capture every aspect of using a model. The source notes that developers' impressions of code quality can differ substantially from benchmark outcomes. The reported cost and score comparison is useful context, but it does not settle how the model will perform in every development workflow.
Weights and access for developers
Deepseek-V4.1-Flash-1017 is available through the company's website, app, and API. Deepseek has also published the model weights on Hugging Face and ModelScope, making the model available as an open-weight release.
Alongside the model, the company has opened its Search and Fetch APIs to developers. Deepseek says the tools use an internally built search engine that handles hundreds of millions of queries per day worldwide. The endpoints cost $0.75 per 1,000 calls.
Deepseek API users can access the features for free through server-side tools in Deepseek Harness, Claude Code, and Codex until October 31, 2026 (4:00 PM UTC). The company points developers to its API documentation for setup details.
What the update signals
The release brings together three changes: a model update focused on coding and agent tasks, open availability of its weights, and developer access to Deepseek's search tools. The company's reported results emphasize a practical tradeoff: a smaller model that can achieve strong scores while using less per-task spending in the cited benchmark.
For teams evaluating it, the published figures offer a starting point rather than a complete verdict. Deepseek's claims still need independent benchmark results, and the source highlights that perceived code quality may not track benchmark scores. The model's availability across Deepseek's own services and model platforms gives developers ways to examine it, while the Search and Fetch APIs add a separate option for incorporating web search into developer tools.