meituan-longcat/LongCat-2.5
AI model · https://huggingface.co/meituan-longcat/LongCat-2.5
First seen by Oriel on .
AI summary
LongCat-2.5 is a 1.6-trillion-parameter native multimodal MoE model designed for long-horizon coding, general agent, visual, creative design, and 3D tasks.
- Approximately 48 billion parameters activate per token, while native-resolution visual features become discrete tokens aligned with text in a shared semantic space.
- Long-horizon agentic RL, multi-teacher distillation, Tree Attention, and long-sequence training target reliable multi-turn execution, reasoning, instruction following, and safety.
- Supernode-aware parallelism, sharding, and communication overlap delivered 86% higher per-device throughput than the DDP baseline on reported long-sequence SFT tasks.
- W4A4KV8 quantization delivered 92% higher decode throughput than W8A8KV16, while W4A4 maintained accuracy comparable to the unquantized model on evaluated benchmarks.
- The MIT-licensed weights support thinking and non-thinking modes, with NPU deployment through SGLang-FluentLLM; developers are advised to assess accuracy, safety, and fairness.
#multimodal models #mixture of experts #agentic coding #model quantization
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Likes
Total likes reported by the source.
- Hugging Face Models: 350, latest observation ; +263 over observed period since
Downloads (30d)
Rolling 30-day downloads. A change in this value is not new downloads during the observed period.
- Hugging Face Models: 1,850, latest observation ; +1,360 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hugging Face Models | meituan-longcat/LongCat-2.5 | https://huggingface.co/meituan-longcat/LongCat-2.5 Read discussion |