DeepSeek has released a new AI model V3.1

# DeepSeek has released a new AI model V3.1

Chinese AI startup DeepSeek has updated its flagship AI model V3 and removed the mention of the reasoning neural network R1 in the chatbot. This is reported by SCMP.

The company announced the release of V3.1 on WeChat. The update expands the model's context window to 128,000 tokens, allowing it to retain more information when interacting with the user. This volume corresponds to a book of approximately 300 pages.

It is also noted that the use of tokens is highly effective.

Data: X. In the Aider Polyglot benchmark, assessing LLM in solving complex programming tasks simultaneously in several languages, DeepSeek V3.1 outperforms Claude 4 Opus.

DeepSeek V3.1 beats Claude 4 Opus on Aider Polyglot

This makes it the best non-TTC coding model and all of that for ~$1 pic.twitter.com/QyJZnVRdVK

— Lisan al Gaib (@scaling01) August 19, 2025

V3.1 strikes a balance between speed and quality of generation. It contains 685 billion parameters and is based on a hybrid architecture that provides high performance in dialogue, reasoning, and programming tasks.

DeepSeek removed the mention of R1 from the deep thinking function. SCMP suggested that this could indicate difficulties in developing the anticipated version R2.

The AI startup DeepSeek attracted attention in January with the release of the R1 model, focused on reasoning. It demonstrated high efficiency with low capital investments, which led to doubts among experts regarding the necessity of billion-dollar inflows into the AI sector and the inflated nature of the industry.

Recall that in June, a Chinese startup began hiring interns to label medical data in order to improve the application of artificial intelligence in hospitals.