# LongCat-2.5 (slim)

> LongCat-2.5 is Meituan's open-weight flagship, a multimodal mixture-of-experts model of 1.6T parameters with about 48B active per token, built for long agentic coding work. MIT licence. The weights are 3,562 GB, and the card documents deployment on NPU hardware only.

- Full: https://www.anchorterminal.com/tools/longcat-2-5.md (~9,350 tokens) · this version ~1,730 tokens · JSON https://www.anchorterminal.com/tools/longcat-2-5.json · canonical https://www.anchorterminal.com/tools/longcat-2-5
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/longcat-2-5#pricing · reviews https://www.anchorterminal.com/tools/longcat-2-5#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 50.5/100 · rank #145 of 160 open-weight models · #107 in Open-weight models · not agent-ready · confidence low**


Assessment: Meituan's flagship is MIT-licensed and documented for tool calling and switchable reasoning, with strong numbers of its own on agentic coding. Almost nobody can run it: 3,562 GB of weights, a 1,336 GB 4-bit build for NPU accelerators, and no released runtime with its architecture. A model to watch until a runtime or a provider picks it up.

## Facts

- Kind: Open-weight model · vendor: Meituan (LongCat) · category: Open-weight models · legal entity: Meituan, named in the card's licence section. The legal entity wasn't looked up · provenance 49/100
- Packages: huggingface `meituan-longcat/LongCat-2.5`, huggingface `meituan-longcat/LongCat-2.5-W4A4`
- Auth: None · pricing: Free · x402: no · licence: MIT
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/meituan-longcat/LongCat-2.5 and the W4A4 repository beside it, read from the model card, the files and their metadata, the config, the chat template, the licence section, the commit history and the runtimes' own code at their latest releases on 10 November 2026. longcat.ai and its blog post are JavaScript pages that gave no text, so nothing on them was read. Meituan's SGLang fork wasn't read. Nothing was downloaded or run
- Model: 1,779,752,648,320 parameters in the files. The card says 1.6 trillion in total with about 48 billion active per token, a native multimodal mixture-of-experts model that turns images into discrete tokens in the same space as text, built on LongCat-2.0 and LongCat-Next
- Files: 210 safetensors files in BF16 (3,562 GB), chat_template.jinja, config and tokenizer files, the LICENSE and figures. No pickle files and no Python. The W4A4 repository holds 1,336 GB with the experts' weights as integers. Hugging Face lists a SHA-256 for the large files
- Licence: MIT, in the metadata, the LICENSE file and the card, which adds that it grants no rights to Meituan's trademarks or patents and that its usage notes don't restrict the licence
- Runtimes: None of the six in a released version. vLLM 0.33.0, SGLang 0.5.23, Transformers 5.22.2 and llama.cpp b12242 have no code for its architecture (LongcatCausalLM, model type pro_v). The card points to a branch of Meituan's SGLang fork for NPU deployment
- Context: The card states no context length. config.json gives 524,288 positions, the tokenizer config says 131,072, and the card's own evaluations ran at 256K, 512K and 1M sequence lengths
- Hardware: No hardware figure on the card beyond NPU platforms. 3,562 GB of weights in BF16 and 1,336 GB at 4 bits
- Training data: The card describes the training in general terms: long-sequence training, reinforcement learning on real engineering tasks and reproduced agent failures, and distillation from several teacher models. No sources, sizes or cut-off. The blog post it points to couldn't be read
- Languages: Not stated. The card says performance varies across languages
- Tool calling: The card's example declares a tool in the OpenAI function format and shows the assistant's tool_calls and the tool message that answers it, with an image earlier in the same conversation
- Reasoning: A hybrid model: enable_thinking on or off per request through chat_template_kwargs, with reasoning returned as reasoning_content. The card recommends keeping earlier reasoning in the history (save_history_reasoning_content) when thinking is on
- Recommended settings: The card's results use thinking on, temperature 1.0 and top_p 0.95 unless a footnote says otherwise

- Scores: Capability pending, Reach 35, Risk 60, Agent interface 75, Documentation and upkeep 46 · total over the 7 assessed categories
- Why:  · Reach, Memory 0, the smallest build is Meituan's W4A4 repository, 1,336 GB, and 1,336 GB × 1.2 = 1,603 GB, far above 256 GB. · Risk, File safety 25, safetensors only, no .bin, .pt or .pth files and no Python in the repository. · Agent interface, Chat template 20, chat_template.jinja in the repository (11,823 characters), with a copy in tokenizer_config.json. · Documentation and upkeep, Model card 11, Intended use, limitations (the usage considerations), the architecture and the parameter counts are on the card.
- Sources: 15, open questions: 7, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/longcat-2-5.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://huggingface.co/meituan-longcat/LongCat-2.5; more in the full twin

## Before you call it

1. You can't run this on a workstation or a single GPU server. Plan for a provider, and none serves it yet
2. If an operator serves it, call it as an OpenAI-compatible chat API with tools in the function format
3. Pass `chat_template_kwargs` with `enable_thinking` true or false
4. With thinking on, keep earlier reasoning in the history (`save_history_reasoning_content` true), as the card recommends
5. The card's results use temperature 1.0 and top_p 0.95
6. Its benchmark table is Meituan's own. The unmarked scores are in-house

## Connect

```bash
# No released runtime loads it. Meituan's NPU instructions are in its SGLang fork:
# github.com/meituan-longcat/SGLang-FluentLLM, branch longcat-2.5-npu, npu_test/ReadMe.md
```

```bash
# The card's example calls an OpenAI-compatible chat completions API with tools, and passes
# extra_body={"chat_template_kwargs": {"enable_thinking": true, "save_history_reasoning_content": true}}
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)