DeepSeek has released DeepSeek-V4-Flash-0731, an updated version of its smaller "Flash" model that surpasses the larger DeepSeek-V4-Pro on independent benchmarks. The release utilizes a new fine-tuning process while maintaining the original Mixture-of-Experts architecture with 284 billion total parameters.
- The model achieves 50 points on Artificial Analysis’ Intelligence Index, tying Gemini 3.6 Flash and trailing only GPT-5.6 Luna and GLM-5.2.
- It costs $0.03 per task on the benchmark compared to $0.05 for GPT-5.6 Luna, placing it on the Pareto frontier for intelligence versus cost.
- Agentic capabilities improved significantly, scoring 1,558 Elo on GDPval-AA and solving 82.7% of problems on Terminal-Bench 2.1.
- Weights are available under the MIT license, with API pricing set at $0.14 per million input tokens.
The update demonstrates that highly intelligent models can be deployed economically for always-on tasks like customer service and bug triaging without requiring proprietary model access.