LongCat-2.0
LongCat-2.0 is Meituan's 1.6-trillion-parameter MoE language model, activating ~48B parameters per token, released under MIT — announced June 30, 2026, with weights and inference code landing on Hugging Face July 5. It is the model behind the stealth "Owl Alpha" that led OpenRouter's charts, and Meituan says pretraining (35T+ tokens, millions of accelerator-days) and serving ran entirely on AI ASIC superpods — accelerators press coverage identifies as Chinese — rather than Nvidia GPUs. LongCat Sparse Attention (LSA) plus training on hundreds of billions of tokens of 1M-context data delivers a claimed 1M-token context window; a 135B-parameter N-gram Embedding module and a 3-step MTP head round out the architecture. Official BF16, FP8, and INT8 releases exist. Meituan reports 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench 2.1, with first-class integration into Claude Code, OpenClaw, and Hermes harnesses.
Positioning
Let's be honest: you are not running this at home. 1.6T parameters means roughly 3.2TB of weights in BF16 and about half that even in the official INT8 release; Meituan's official SGLang recipe deploys it across 16 H20, H200, or B200 GPUs spanning two nodes, and even a hypothetical 2-bit quant would exceed 400GB before KV cache. No consumer or workstation path exists — "local" here means a multi-node cluster, full stop.
Who should run it: enterprises with datacenter GPU or NPU capacity that want a near-frontier agentic coder under a genuinely clean MIT license — no acceptable-use policy, sublicensing and resale allowed — for data-sovereignty or air-gapped deployments. Researchers get a rare artifact too: a trillion-scale MoE with published sparse-attention (LSA) and N-gram embedding designs, trained entirely off Nvidia silicon.
Honest tradeoffs: the numbers are near-frontier, not frontier — SWE-bench Pro 59.5 edges GPT-5.5's cited 58.6 but sits well behind Claude Opus 4.8's 69.2, and BrowseComp 79.9 trails the leaders. All headline scores come from Meituan's own harness. For everyone without a cluster, the practical way to use LongCat-2.0 is a hosted API — which defeats the point of a local-AI catalog. If you want an open agentic coder you can actually rack, Hy3 at 295B or GLM-class models are the realistic ceiling; LongCat-2.0 is here because its license and provenance matter, not because you can run it.
Overview
LongCat-2.0 is Meituan's 1.6-trillion-parameter MoE language model, activating ~48B parameters per token, released under MIT — announced June 30, 2026, with weights and inference code landing on Hugging Face July 5. It is the model behind the stealth "Owl Alpha" that led OpenRouter's charts, and Meituan says pretraining (35T+ tokens, millions of accelerator-days) and serving ran entirely on AI ASIC superpods — accelerators press coverage identifies as Chinese — rather than Nvidia GPUs. LongCat Sparse Attention (LSA) plus training on hundreds of billions of tokens of 1M-context data delivers a claimed 1M-token context window; a 135B-parameter N-gram Embedding module and a 3-step MTP head round out the architecture. Official BF16, FP8, and INT8 releases exist. Meituan reports 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench 2.1, with first-class integration into Claude Code, OpenClaw, and Hermes harnesses.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
Weaknesses
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of LongCat-2.0.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
Can I use LongCat-2.0 commercially?
What's the context length of LongCat-2.0?
Source: huggingface.co/meituan-longcat/LongCat-2.0
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify LongCat-2.0 runs on your specific hardware before committing money.