In this video, I'll be telling you about Meituan’s new LongCat-2.0 model, a massive 1.6 trillion parameter Mixture-of-Experts model that could become one of the biggest open-weight AI models once the weights are fully uploaded.

--
Key Takeaways:

🚀 Meituan has released LongCat-2.0, a huge Mixture-of-Experts model with 1.6 trillion total parameters.
📦 The Hugging Face page is live, but the actual model weights are still being uploaded.
🧠 LongCat-2.0 activates around 48 billion parameters per token, making it much larger than the first LongCat model.
📚 The model includes LongCat Sparse Attention and was trained for long-context, coding, research, and agentic tasks.
⚙️ Meituan says the model was trained and deployed on AI ASIC superpods instead of the usual Nvidia GPU setup.
📊 Official benchmarks look strong, especially for agentic coding, SWE-bench, reasoning, and long-context tasks.
🧪 In my one-shot testing through the free chat site, LongCat-2.0 did not perform very well on coding tasks.
🤖 The model may perform better in a real agent setup, but the API and coding plan are currently only available in China.
🌍 Even with mixed test results, a 1.6 trillion parameter open-weight model is still a major release for the open-source AI ecosystem.
👍 Overall, LongCat-2.0 is an exciting open model release, but I need proper agent testing before judging its real coding ability.

AICodeKing

Meituan released LongCat 2.0, a massive 1.6 trillion parameter mixture-of-experts open model with ~48B active parameters per token. It features LongCat sparse attention for efficient long-context processing, an engram embedding module, and was trained on over 35 trillion tokens using non-NVIDIA AISX super pods — a notable statement about frontier-scale training on alternative hardware. The model targets agentic coding and long-context tasks. However, weights are not yet available on Hugging Face, and the API/coding plan is currently China-only. Testing via the free chat site yielded poor one-shot results (21.6% on KingBench 3), underperforming most major models. The reviewer suspects the model may be optimized for agentic workflows rather than one-shot tasks, and reserves final judgment until proper agent-based testing is possible once weights or API access become available globally.

Meituan LongCat 2.0 (Tested): China's 1.6T OPEN MODEL looks CRAZY!

<p>This is a fascinating case study in how hard it is to evaluate frontier models. LongCat 2.0 represents a genuine milestone—1.6T MoE parameters trained on alternative hardware (non-NVIDIA AISX super pods) is a statement about both technical capability and supply chain diversification. The 21.6% on KingBench 3 is disappointing on the surface, but the reviewer’s suspicion that it’s optimized for <em>agentic</em> workflows rather than one-shot tasks is a crucial nuance. Many agent-optimized models perform poorly on static benchmarks because they’re tuned for multi-turn, tool-using, iterative reasoning—not single-shot Q&amp;A. The real test will be when the weights are available globally and we can run proper agent-based evals. Until then, reserving judgment is the right call. The technology is impressive even if the early benchmarks aren’t.</p>
