DeepSeek, a Chinese company, released a new AI model, "DeepSeek-V4.1-Flash," on 2026-09-10. The model supports image input in addition to text.

DeepSeek-V4.1-Flash is a Mixture of Experts (MoE) model with a total of 552 billion parameters, of which 8 billion are active during input and 16 billion are active during output. The company stated that the adoption of proprietary pre-training and reinforcement learning methods has achieved both higher performance and increased speed.

In multiple benchmark tests, DeepSeek-V4.1-Flash demonstrated performance surpassing Claude Opus 5 and GPT-5.6 Sol. In the "Artificial Analysis Intelligence Index v4.3," a comprehensive performance metric by Artificial Analysis, it recorded a score higher than GPT-5.6 Luna. Furthermore, it was reported to have achieved the top score on "AutomationBench-AA," which measures agent performance for tasks such as office work, surpassing GPT-6 Astra.

Due to improvements in inference, API pricing has been set at a low cost. Rates vary by time of day, with input costs during peak hours at $0.3 per 1 dollar, and half that price during off-peak hours. Additionally, the model is designed to significantly reduce memory and storage consumption through a KV cache mechanism.

The model has been released free of charge under the MIT License via Hugging Face and ModelScope.

Source: