Meituan releases LongCat-2.5, a 1.6T multimodal agent model
Original title: meituan-longcat/LongCat-2.5 — new model trending #30 on Hugging Face
The mixture-of-experts model activates about 48 billion parameters per token for coding, visual, and agent tasks.
Meituan LongCat published LongCat-2.5, a native multimodal mixture-of-experts model for long-horizon agentic work, building on LongCat-2.0. The card lists 1.6 trillion total parameters and about 48 billion activated per token, with visual features converted to discrete tokens and aligned with text. In-house benchmark scores are described as indicative, while starred comparison scores are cited from other reports.
Unconfirmed: Starred comparison scores are cited from other reports, and unmarked scores are in-house and indicative only.
- The model card lists an MIT license and Hugging Face and ModelScope links.
- Terminal Bench v2.1 is scored 88.4 in-house, versus 70.8 for LongCat-2.0.
- Long-sequence SFT is reported at 86% higher per-device throughput than the DDP baseline.
- KV-cache changes nearly double decode KV capacity without raising TPOT on the tested workload.
- Organizations
- Meituan LongCat
- AI models
- LongCat-2.5LongCat-2.0
From the original article
6 min read at huggingface.coimage-text-to-text · 971 downloads · 237 likes · license mit
By meituan-longcat · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.