What is LongCat-2.5?
LongCat-2.5 is a mixture-of-experts model from Meituan with 1.6 trillion parameters in total and about 48 billion active per token. It reads images and video natively, turning visual features into tokens in the same space as text, and is trained for long-horizon agent work: whole-project coding builds, browsing, research and 3D tasks.
Meituan reports results across more than 30 benchmarks against GPT, Claude, GLM, Kimi and DeepSeek models, including 64.6 on DeepSWE v1.1, 91.9 on BrowseComp and 83.1 on MCP Atlas. The weights are MIT-licensed, and Meituan documents deployment on NPU accelerators through its SGLang fork.
Is LongCat-2.5 worth it?
LongCat-2.5 is Meituan's bid for the top of the open-weights table on agent work. It is a 1.6-trillion-parameter mixture-of-experts model with about 48 billion active parameters per token, trained with large-scale reinforcement learning on real engineering tasks and distilled from specialist teacher models for agents, reasoning and conversation.
The reported numbers put it in frontier company. Meituan lists 64.6 on DeepSWE v1.1, 91.9 on BrowseComp, 83.1 on MCP Atlas and 63.1 on Humanity's Last Exam with tools, close to or above several closed models in its own table. It also reads images and video natively, scoring 85.2 on VideoMME, and handles visual coding tasks such as text-to-3D. The MIT licence is the most permissive of the large open models.
The limits are practical and evidential. Serving 1.6 trillion parameters needs a large accelerator cluster; the published deployment guide targets NPUs, and GPU serving details are thin on the model card. Most comparison scores are marked as taken from other reports, several benchmarks are Meituan's own, and agent results depend on harness settings such as Claude Code with 512K-token sequences. Independent replication will matter more than usual.
Where it earns its place
An enterprise AI platform team with an accelerator cluster
Host an open agentic coding model with no licence restrictions
Deploy LongCat-2.5 through SGLang-FluentLLM and put it behind your coding agents; the MIT licence allows internal and commercial use without territorial limits.
A research group studying agent training
Compare against a strong open baseline
Use the published benchmark settings and harnesses to replicate scores before building on them.
Before you commit
- Setup
- Download the weights from Hugging Face and deploy with Meituan's SGLang-FluentLLM; the documented path is for NPU platforms.
- Integrations
- SGLang-FluentLLM; evaluated inside Claude Code and mini-swe-agent harnesses.
- Data and privacy
- Self-hosted; MIT licence.
- Cost at scale
- No licence fee; the cost is a multi-node accelerator cluster to hold 1.6T parameters.
- Limits
- Hardware requirements are not published as a single figure; several results depend on long contexts of 256K to 1M tokens.
What’s good
- Strong reported agentic coding and browsing scores
- Native image and video understanding
- MIT licence
- About 48B active parameters per token for its size
What’s not
- 1.6T parameters need a large cluster
- Deployment guide targets NPUs; GPU details thin
- Several benchmarks are Meituan's own
- Comparison scores partly taken from other reports
Best for
Labs and large companies with serious inference hardware that want an open, permissively licensed agentic model to host themselves.
LongCat-2.5 alternatives
LongCat-2.5: common questions
- Is LongCat-2.5 worth it?
- LongCat-2.5 is among the strongest open agentic models on paper, under a licence with almost no strings, but at 1.6 trillion parameters it is a model for clusters, not companies with one GPU server. We rate it 4.1 out of 5.
- How much does LongCat-2.5 cost?
- Free open weights under the MIT licence; running it needs a large multi-accelerator cluster.
- What are the downsides of LongCat-2.5?
- 1.6T parameters need a large cluster; deployment guide targets NPUs; GPU details thin; several benchmarks are Meituan's own; comparison scores partly taken from other reports.
- Who is LongCat-2.5 best for?
- Labs and large companies with serious inference hardware that want an open, permissively licensed agentic model to host themselves.
- How much does LongCat-2.5 cost for a team or in production?
- No licence fee; the cost is a multi-node accelerator cluster to hold 1.6T parameters.
- Is LongCat-2.5 safe for company data?
- Self-hosted; MIT licence.
- What are the best LongCat-2.5 alternatives?
- The closest alternatives we have reviewed are LongCat-2.0.
Answers checked Nov 15, 2026.