Skip to content
Neural Daily

ModelsNews

LongCat-2.5: a 1.6T open agent model under MIT

🇭🇺 Olvasd el magyarul →

Meituan published the LongCat-2.5 weights on Hugging Face. The natively multimodal MoE model has 1.6 trillion parameters, about 48 billion active per token, under the MIT license. Meituan's numbers put it near closed frontier models on agent tasks, though not everywhere.

· 3 min read

A Meituan LongCat modellcsalád logója
Image: Meituan LongCat · Press kit · Source

What Meituan released

LongCat-2.5 is the new model from Meituan's LongCat team and the successor to LongCat-2.0. It targets long, multi-step agent work: large codebase development, full-project builds, visual design and 3D tasks. The weights are on Hugging Face under the MIT license (Hugging Face).

  • Size: 1.6 trillion total parameters, about 48 billion active per token (MoE).
  • Natively multimodal: images become discrete tokens aligned with text in a shared space.
  • Hybrid mode: thinking can be switched on or off, with off being more token-efficient.
  • OpenAI-compatible calls with tool calling, with sample code on the model card.

What the numbers show

The model card compares LongCat-2.5 with Claude Opus 5, GPT-5.6-Sol, GLM-5.3, Kimi K3 and DeepSeek-V4.1-Flash:

  • Terminal Bench v2.1: 88.4. Claude Opus 5: 89.1, GPT-5.6-Sol: 88.8, DeepSeek-V4.1-Flash: 90.6.
  • BrowseComp: 91.9, the best in the table (Opus 5: 90.8).
  • AutomationBench: 51.8. Opus 5: 50.3, DeepSeek-V4.1-Flash: 54.8.
  • Terminal Bench v4.0: 38.5, where Opus 5 leads clearly with 53.9.
  • NL2Repo: 60.5, against 75.3 for Opus 5.
A LongCat-2.5 benchmarkeredményei kódolási, általános agent és vizuális feladatokon a Claude és GPT modellekkel összevetve
Meituan LongCat · Screenshot · Source

Source: Meituan LongCat, Hugging Face model card

Two caveats. Scores without an asterisk are Meituan's in-house measurements, which the card calls indicative only, and there is no independent evaluation yet. Also, the table compares against Opus 5 and GPT-5.6-Sol, not the newest closed models.

Quantization and serving

Expert weights and activations are quantized to W4A4 for native INT4 compute. Meituan reports 43.3% higher prefill and 42.77% higher decode throughput than a W8A8 MoE baseline. With KV cache quantization (W4A4KV8), decode throughput is 92% higher than W8A8KV16.

In practice this is not a single-GPU model. The card points to an NPU deployment guide built on SGLang-FluentLLM, and no other inference stack is documented yet. You can try it without downloading at longcat.ai.

Why it matters

An MIT license is rare at this size: commercial use and fine-tuning are allowed, with no rights granted to Meituan trademarks or patents. That puts LongCat-2.5 among the leading open-weight agent models alongside Kimi K3.1 Pro (Kimi K3.1 Pro: Moonshot's open-weight model nears Opus). If you plan to self-host agents, test it on your own workloads, especially long, hard tasks like Terminal Bench 4.0, where the gap is still wide.

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles