DeepSeek releases V3.1 with hybrid reasoning architecture
DeepSeek-V3.1 released August 21, combining thinking and non-thinking modes with 128K context, with API price hike from September 6.
Two weeks after OpenAI launched GPT-5, DeepSeek officially released DeepSeek-V3.1 on August 21. The core highlight is a hybrid reasoning architecture that supports both thinking and non-thinking modes within a single framework, effectively merging DeepSeek-R1 and DeepSeek-V3.
V3.1 uses the UE8M0 FP8 Scale parameter precision designed for next-generation domestic chips. The API context window expanded to 128K, with added support for the Anthropic API format to simplify migration. The official app and web client were upgraded simultaneously, letting users toggle thinking mode via a dedicated button.
Also announced was an API price adjustment: from 12:00 AM Beijing time on September 6, input price for cache-miss rises to ¥4 per 1M tokens (V3 was ¥2), output adjusts to ¥12 per 1M tokens (previously ¥8), and off-peak discounts are cancelled. This signaled a shift from price wars to value competition among Chinese LLM vendors.