DeepSeek v3.1 released with hybrid inference, faster thinking, and upgraded agent skills
DeepSeek v3.1 has been released as the latest version of this Chinese large language model. This update introduces hybrid inference, letting users switch between Think and Non-Think modes using the new DeepThink button for seamless mode selection.
Building on this, DeepSeek-V3.1-Think now delivers answers in less time compared to DeepSeek-R1-0528. Users who depend on rapid output will notice faster response times during inference.
Alongside performance improvements, post-training updates strengthen the model’s agent-related skills. Tool use is more reliable, and multi-step agent task execution is enhanced. These changes are aimed at supporting more complex workflows involving tool invocation and sequential processes.
Developers and integrators also benefit from a broader set of upgrades. This release includes updated APIs, expanded toolsets, and improvements for agent-related components. Additional model updates further extend DeepSeek-V3.1’s capabilities across multiple usage scenarios.
- AI Chatbot
- Freemium • Open Source
Related news
External links
- DeepSeek-V3.1 ReleaseDeepSeek • Official source
- DeepSeek's new V3.1 hints at potent new Chinese chipsThe Register
- Chinese AI startup DeepSeek releases upgraded model with domestic chip supportReuters
- DeepSeek Releases V3.1 Model: What’s New?Gizmochina
Comments
Given how little cards China has access (compared to US), and how outdated theses Nvidia H20 are, the benchmarks are pretty good. This also avoid steep API price raises, keeping the price very sensible. And the minor version number increase is way more honest than bumping to the next major version every few months like some others are doing, when obviously, it's pretty much always the same underlying model.
Releasing huge improvements with small version numbers comes across as being humble about it, at least to me.