Overview

• DeepSeek says V3.1 operates with “one model, two modes,” letting users toggle between thinking and non‑thinking for quicker answers.
• The update extends context to 128,000 tokens and introduces stronger tool use and function calling with OpenAI- and Anthropic‑compatible APIs.
• The base model is live on Hugging Face, with reports citing roughly 671–685 billion parameters and an Aider coding score of 71.6% at very low per‑task cost.
• DeepSeek announced revised V3 pricing starting in early September, including higher fees for some services and removal of discounted evening rates.
• References to the R1 reasoner were removed from the chatbot, prompting reports that an R2 successor may be delayed, as Baidu’s open‑sourced Ernie intensifies competitive pressure.