DeepSeek-v3.1 (api-docs.deepseek.com)

Article summary

DeepSeek-V3.1 is a hybrid inference model that combines thinking and non-thinking modes, with improvements in thinking efficiency, tool use, and multi-step agent tasks. The model has been updated with a new tokenizer and chat template, and its open-source weights are available. The API has also been updated with new features, including strict function calling and support for the Anthropic API format. Pricing changes have been announced, with new rates starting on September 5th, 2025.

Main themes

• Hybrid Inference Models
• AI Model Updates
• API Development
• Pricing and Cost
• Benchmarking and Evaluation
• Model Comparison

What commenters say

• The performance of DeepSeek-V3.1 is reasonable compared to other open-weight models, but it does not surpass top models like GPT-5 and Claude 4.
• Some users have reported positive experiences with DeepSeek-V3.1, citing its ability to produce high-quality results and avoid hallucinations.
• The use of benchmarks to evaluate AI models is problematic, as companies may be able to cheat by tuning their models to perform well on specific benchmarks.
• The secrecy of benchmarks is not effective, as companies can still access and use the data to improve their models.
• The pricing of AI models like DeepSeek-V3.1 is a significant concern, with some users feeling that it is too expensive.
• The format of tool calls in DeepSeek-V3.1 can be inconsistent and may require additional support to accommodate.
• Some users believe that the only way to truly evaluate an AI model is to test it oneself, rather than relying on benchmarks or other evaluations.
• The training data and methods used to develop AI models can have a significant impact on their performance and behavior.