DeepSeek released the DeepSeek-V4.1-Flash-1017 update on Friday, October 17, 2026. The model, an evolution of the September version, improves performance in agent tasks, coding and reasoning benchmarks, with shorter outputs and faster execution, while keeping the same API price.
What is DeepSeek V4.1-Flash-1017?
It is a Mixture-of-Experts model with 552 billion parameters, Causal Encoder-Decoder architecture, 1 million token context and native multimodal support. Weights are open under MIT license on Hugging Face and ModelScope. On the API, it remains accessible via the deepseek-flash identifier with no price changes.
According to the company, the update represents a Pareto improvement: better scores with lower cost per task. Examples include DeepSWE v1.1 from 74.2 to 77.0, Terminal-Bench 4.0 from 31.2 to 42.3 and gains in ProgramBench, Automation-Bench and others. In some agent tests, it outperforms larger models like Claude Opus 5 and GPT-5.6 Sol in specific categories.
New tools for agents
Alongside the model, DeepSeek opened its Search and Fetch APIs to all developers, based on an internal search engine. Each endpoint costs US$ 0.75 per thousand calls, with a free period until October 31, 2026 for certain uses.
The update is already live on DeepSeek's website, app and API. Sources: DeepSeek API Docs, Hugging Face, The Decoder.
By GeekikiBot