DeepSeek releases powerful open-source V4.1-Flash-1017 model

DeepSeek has released an updated version of the DeepSeek-V4.1-Flash-1017 model, which significantly outperforms previous versions in agentic tasks.

The model features 763 billion parameters and supports a context window of up to one million tokens. It can analyze images, invoke external tools, and operate in reasoning mode.

Unlike closed-source alternatives, the model is distributed under the MIT license. This allows users to freely download the weights and run them on their own hardware, provided sufficient computational resources are available.

This update demonstrates substantial performance gains compared to the base V4.1-Flash version.

• Terminal-Bench 4.0: score increased from 31.2 to 42.3
• ProgramBench: result improved from 20.3 to 29.1
• Apex Bench: metric rose from 64.6 to 72.1

In several tests, the new model surpassed closed solutions like Opus 5 and GPT-5.6 Sol. In the DeepSWE task, it scored 77.0 points compared to 74.0 for Opus 5. The margin in Automation-Bench was 60.5 versus 50.3.

However, competitors still maintain leadership in complex, multi-step tasks. In Humanity's Last Exam, Opus 5 achieved 56.3 points against DeepSeek's 42.6.

Developers also note that the model executes tasks faster and generates more concise responses. The model is available for download on the Hugging Face platform.