DeepSeek V4.1 Flash scores 90.6 on Terminal-Bench 2.1, beating GPT-5.6 Sol
The Chinese AI developer launched the V4.1 Flash model, which uses a specialized architecture to activate only a fraction of its total parameters for faster and cheaper inference. Its performance on coding and cybersecurity benchmarks surpassed competitors including OpenAI's latest models.
DeepSeek announced the release of its new V4.1 Flash artificial intelligence model. The company stated that this version outperforms its previous flagship and rivals competitor Kimi K3. The developer utilized a specialized Causal-Encoder-Decoder architecture based on a massive 552 billion-parameter framework. However, to reduce costs, it employs a Mixture-of-Experts design that routes tasks only to relevant subnetworks. This specific configuration activates roughly 8 billion parameters for processing inputs and 16 billion for generating responses, significantly lowering computing power needs per request. The model supports native multimodal visual understanding and remains the smallest in its new series. DeepSeek reported that V4.1 Flash achieved a score of 90.6 on Terminal-Bench 2.1. OpenAI's GPT-5.6 Sol scored 88.8 on the same test, which evaluates real-world compute tasks. Moonshot AI's Kimi also performed on this benchmark. The release occurs as Chinese firms compete against global tech giants to commercialize AI. Rising hardware costs and foreign chip export restrictions have tightened computational constraints in China.
