DeepSeek quietly released its latest large language model, V3.1, on August 19, 2025, introducing a hybrid reasoning architecture that integrates conversational and reasoning capabilities into a single model. The base model was made available on Hugging Face without official marketing promotion.
- DeepSeek V3.1 achieves a 71.6% pass rate in Aider programming tests, slightly surpassing Claude Opus's 70.6%.
- The model costs approximately $1 for the tested suite of 225 tasks, offering a 68x cost advantage over Claude Opus.
- It features a 685B parameter architecture with an extended context window of 128k tokens, up from 64k in previous versions.
- The hybrid design automatically selects reasoning depth to reduce unnecessary computation while maintaining high performance.
The release highlights DeepSeek's strategy of letting product performance speak for itself, providing enterprises with a cost-effective alternative for code generation and debugging.