AI models 1 min read

LongCat-Flash-Chat vs DeepSeek V3: variable active parameters or a fixed 37B

A non-thinking 560B Mixture-of-Experts chat model with dynamic computation against DeepSeek V3 with 37B activated per token.

How active parameters are described

LongCat-Flash-Chat is a non-thinking chat model from LongCat-Flash, a Mixture-of-Experts model with 560 billion total parameters. Its README says a zero-computation experts mechanism activates between 18.6 and 31.3 billion parameters depending on the token, about 27 billion on average.

DeepSeek V3's README states 671B total parameters with 37B activated for each token and lists a 128K context length. The LongCat-Flash README says mid-training extended the context length to 128k.

Licences and deployment

The LongCat-Flash-Chat README says the model weights are released under the MIT License and that the licence grants no rights to Meituan trademarks or patents. DeepSeek V3's README says its code repository is licensed under the MIT License, while use of the Base and Chat models is subject to the Model License.

LongCat-Flash-Chat lists basic adaptations in SGLang and vLLM for deployment, with a separate deployment guide.

Who should look at which

Readers who want a model whose compute per token varies and whose weights are stated to be under the MIT License may start with LongCat-Flash-Chat.

Readers who want a fixed 37B activated per token and a documented 128K context should start with DeepSeek V3 and read its Model License. Nothing was installed, run or benchmarked, and the benchmark tables in both READMEs are the publishers' own.

VendorClue read the project pages for LongCat-Flash-Chat on 7 November 2026 and DeepSeek V3 on 4 November 2026. Terms and releases can change, so check the current pages before you choose.

Sources

LongCat-Flash-Chat: longcat-flash-chat-project · Evidence: documentation · Not checked

DeepSeek V3: deepseek-source · Evidence: documentation · Not checked

Next step

Tools in this guide

All AI models tools
All guides