🐈だけど、強い！

【LongCat-Flash-Thinking】 562BでMoEアーキテクチャのモデルが登場。推論時は18.6B∼31.3Bが有効になるそう。 ベンチマークではGPT-5-Thinkingに肉薄しています。 ライセンスはMIT！

Up on chutes, including function calling and reasoning parsing! Same deal as the other optional thinking models, set "chat_template_kwargs": {"enable_thinking": true/false}, or X-Enable-Thinking: true request header, or tack on ":THINKING" to model name. (Thinking is enabled by default it seems) Also remember that because it uses a reasoning parser, the response body/SSE chunks have reasoning_content and content response fields, separated. Hacks: github.com/jondurbin/sglang/… github.com/rayonlabs/chutes-…

美团LongCat这么猛，这是什么性能怪兽？让我接入 Claude Code 试试看。

Not a Chinese bot glorifying the Chinese Communist Party lmfao.

SWE Bench score?

@chutes_ai you guys should add this one. Looks Goated.

We've got delivery companies in China building better LLMs than OpenAI for like probably 100 bucks

what the helly this looks good

this is probably the most interesting part from the linked paper

These are really good benchmark numbers 👏

Man that cat is long. and it is very good cat. love it. it is SOTA model for coding, agentic, and Logic tasks!

These guys COOKED > On AIME25, it achieves high accuracy with 64.5% fewer tokens, thanks to native tool use + agent-first design > Asynchronous RL framework offers a 3× speedup over synchronous workflows.

Still bizarre how Meituan is making LLMs, and it's not a run-of-the-mill recipe, they legit have done serious research here, improving on V3 architecture. Not sure about benchmarks, learning good data methods might take more time. But bizarre and fascinating.

Meituan cooked

For all those thinking the AI race is going to be decided between US companies only: Meituan, the Chinese version of DoorDash, casually drops a super fast version of LongCat which rivals OAIs GPT-5. Great effort by the LongCat team! Cooked up an amazing model!

Another Meituan release! They deliver quickly. 🚀 LongCat-Flash-Thinking is out—designed for smarter reasoning at lower costs. 🏆 Benchmarks show top performance in logic, math, coding, and agent tasks. 📊 On AIME25, it achieves high accuracy with 64.5% fewer tokens, thanks to native tool use and agent-friendly design. ⚙️ Its asynchronous RL framework offers a 3× speedup over traditional synchronous setups. 🔗 Model: huggingface.co/meituan-longc…… 💻 Try it: longcat.ai

a lot of alpha in this technical report , the policy optimizations and the @Kimi_Moonshot -like RL environment are a must read link : github.com/meituan-longcat/L… their DORA RL environment is perfectly described here (and more) 👇🏻

LogCat Flash Thinking outperforming - GPT-5 on AIME-24 - Gemini 2.5 Pro & o3 on ARC-AGI - o3 & Qwen3-225B-2507 on LiveCodeBench was not on my bingo card today! MIT licensed and so fast... 😳 What a release!🔥🔥🔥 Meituan is now officially a frontier lab!