"DeepSeek-V4-Flash-0731" debuts: an open model with comparable performance to GPT-5.6 Luna at a lower price

DeepSeek released DeepSeek-V4-Flash-0731 under a commercial-friendly MIT license and published it on Hugging Face and ModelScope.

The model keeps the same architecture as the preview version and uses an MoE design with 284B total parameters and 13B active parameters.

Third-party benchmarks place it on par with Gemini 3.6 Flash and ahead of GPT-5.6 Luna on agent tasks.

It is also cheaper to run than GPT-5.6 Luna after the late-July price cut, strengthening competition among high-performance open models.