DeepSeek-V4-Flash—a new version of the lightweight, fast model—is available in public beta via the API. This is no ordinary update; the release focuses primarily on enhancing agent capabilities.
The model can now function like an engineer: opening a terminal, analyzing repositories, invoking tools, and executing complex, multi-step tasks. DeepSeek-V4-Flash is becoming a full-fledged AI assistant for developers and autonomous agents.
What’s new in the numbers?
Benchmark results:
- Terminal Bench 2.1 (terminal operations): 82.7
- Cybergym (cybersecurity): 76.7
- Toolathlon Verified (tool challenge): 70.3
- DSBench-FullStack (full-stack development): 68.7
- Terminal Bench 2.1 (terminal operations): 82.7
- Cybergym (cybersecurity): 76.7
- Toolathlon Verified (tool challenge): 70.3
- DSBench-FullStack (full-stack development): 68.7
Technical details
Structure: the architecture and size remain unchanged (13 billion active parameters in total), but the model has undergone full retraining (post-training).
Where it applies: the update affects only the API version of deepseek-v4-flash. The models on the website (WEB) and in the app (APP) remain unchanged.
Integration: the model supports the Responses API format and is optimized for use with Codex.
What's next: the release of the full version of DeepSeek-V4-Pro is expected soon.
0 Comments