What is DeepSeek-V3.1, and what does a hybrid thinking model mean?

Yesterday, August 21, 2025, the Chinese AI company DeepSeek released DeepSeek-V3.1, an update to the model family that drew worldwide attention earlier this year. DeepSeek calls it "our first step toward the agent era." The headline change is that one model now does two jobs: quick everyday answers and slower, step-by-step reasoning.

Until now, DeepSeek users picked between two separate models: V3 for chatting and R1 for hard reasoning. V3.1 folds both into a single system with two modes, which DeepSeek calls Think and Non-Think.

What DeepSeek released

A context window is how much text the model can consider at once; 128,000 tokens (a token is a word or piece of a word) is roughly a long novel. Open weights means the trained model itself is published, so anyone with enough computing power can run it or adapt it.

How a hybrid thinking model works

A reasoning model is trained to write out intermediate steps, a kind of scratchpad, before giving its final answer. That helps on math, logic and coding, but it costs time and tokens. For simple requests, like rewording an email, the extra thinking adds delay without improving the answer.

A hybrid model learns both behaviors and lets you choose. Think of a colleague who can either answer off the top of their head or say "let me work through this on paper" when the question deserves it. The same knowledge sits behind both; only the working style changes.

V3.1 is built on top of DeepSeek-V3, the large mixture-of-experts model released last December. That design has 671 billion parameters in total, but only 37 billion are active for each token: each piece of text is handled by a small set of specialist sub-networks rather than the whole model. For V3.1, DeepSeek continued training the base model on another 840 billion tokens to extend how much text it can handle, updated its tokenizer (the tool that splits text into tokens), and then trained it further to use tools and carry out multi-step agent tasks.

Here is a simple worked example. Ask the same question in each mode:

A shop sells pens only in packs: 3 pens for $2 or 10 pens for $6.
What is the cheapest way to get at least 21 pens?

In Non-Think mode, you should get a fast, short answer. In Think mode, you will see the model work through the options first: seven small packs give 21 pens for $14; one big pack and four small ones give 22 pens for $14; two big packs and one small one give 23 pens for $14. The minimum is $14 either way, and the careful version is more likely to notice that the same $14 can buy 23 pens.

The chip angle

In a comment on its official WeChat account, DeepSeek said V3.1 uses a precision format called UE8M0 FP8, which it described as tailored for next-generation chips made in China that will be released soon. FP8 means storing each number in 8 bits instead of 16 or 32. Fewer bits per number is a bit like rounding prices to the nearest dollar: you lose a little detail but can handle far more numbers with the same memory and computing power.

CNBC notes that DeepSeek did not say which chips it trained V3.1 on or which Chinese chips the format targets. The comment matters because of the wider context: the United States restricts exports of advanced AI chips to China, and Beijing has reportedly been pressing the country's AI developers to replace Nvidia's processors with Chinese-made ones.

Why it matters this week

For users, V3.1 simplifies the choice: one model, with a button for when a question deserves more effort. For developers, the improved tool use and Anthropic-compatible API make it easier to plug DeepSeek into agent tools they already use, and the open weights let companies run it on their own servers.

For the industry, it is another sign that the line between chat models and reasoning models is fading, and that capable models keep arriving as free downloads. If DeepSeek's chip hint proves out, future models may also show how far Chinese hardware can carry advanced AI.

How do you choose between thinking and non-thinking mode in DeepSeek-V3.1?

On chat.deepseek.com, leave DeepThink off for everyday tasks: summaries, translations, rewrites, quick facts and brainstorming. Turn it on for problems with several steps where mistakes are easy: math, logic puzzles, debugging code, comparing options with many conditions, or planning a project. If an answer seems too quick or misses part of the question, ask again with DeepThink on.

Developers make the same choice through the model name: deepseek-chat for fast, cheaper replies and deepseek-reasoner when accuracy matters more than speed. Check DeepSeek's pricing page before September 5, when the new rates begin.

In either mode, verify important answers. Thinking mode makes errors less likely, not impossible, and as with any online AI service, think about what personal or business information you share in your prompts.