Deepseek has released the V4.1-Flash-1017 model, which improves performance in coding and agentic tasks compared to its predecessor. The model is available via the company's website, app, and API, with weights published on Hugging Face and ModelScope. It utilizes reinforcement learning with self-feedback to generate software projects, 3D models, and animations without human evaluators.
In internal benchmarks, the model outperformed larger competitors like Claude Opus 5, GPT 5.6-Sol, Kimi K3, and GLM 5.3. Efficiency gains are significant: the new model achieves a 77 percent score at a cost of under $0.20 per task, while the previous version cost around $0.50. Competing models from other providers currently range from $2 to over $10 per task.
The company is also opening its proprietary search engine to external developers via Search and Fetch APIs. These services handle hundreds of millions of queries daily. The cost is set at $0.75 per 1,000 calls, though free access is available through specific developer tools until October 31, 2026.
Future availability of such models may be affected by potential new Chinese government regulations. Reports suggest authorities are considering stricter rules or bans on the release of open-weight models, particularly those ranked among the top 10 most capable systems.
