Meituan releases LongCat-2.5, a 1.6 trillion parameter multimodal model for long agentic tasks
The model activates about 48 billion parameters per token, offers a one-million-token context and is available through an API; Meituan's claims of parity with Claude Opus models on several benchmarks are self-reported.
Why it matters
LongCat-2.5 adds to a run of very large Chinese models aimed at agentic coding, and Meituan says on its site that training was done entirely on domestic chips, which matters under US export controls. It builds on LongCat-2.0 from June and a September preview that came without benchmark results. Watch for independent evaluations, whether open weights follow on Hugging Face, where earlier LongCat models were published, and how its pricing compares with rival Chinese models.
Key facts
- 1.6 trillion total parameters, about 48 billion active
- One-million-token context, up to 128,000 output tokens
- Released November 2, 2026, via API
- Builds on LongCat-2.0 from June 2026
- W4A4 quantization for inference
- Training done on domestic chips, Meituan says