What is LongCat-2.0?
LongCat-2.0 is Meituan's largest open-weight model. It surfaced on OpenRouter under the codename Owl Alpha, was named on 30 June 2026, and had its weights published on Hugging Face in early July in BF16, FP8 and INT8. It is a mixture-of-experts model with 1.6 trillion parameters and about 48 billion active per token, including 135 billion parameters of N-gram embeddings, and a 1M-token context. It is text only, with thinking that can be switched on or off.
Meituan says it was pretrained on more than 35 trillion tokens entirely on AI ASIC superpods rather than NVIDIA GPUs, and it can be served on GPUs or NPUs. The weights are under the MIT licence. On Hugging Face it has 285 likes and 2,957 downloads in the last 30 days, a small number that reflects how few teams can host it.
On OpenRouter, AtlasCloud serves it at $0.30 per million input tokens and $1.20 per million output tokens, with cached input at $0.006 and up to 262,144 output tokens. Meituan has since launched LongCat-2.5 on its own API, but has not released its weights.
How much does LongCat-2.0 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | MIT licence |
| OpenRouter (AtlasCloud) | $0.30 in / $1.20 out per 1M tokens | $0.006 cached input |
Open weights under the MIT licence. OpenRouter (AtlasCloud): $0.30 input / $0.006 cached / $1.20 output per 1M tokens.
Is LongCat-2.0 worth it?
LongCat-2.0 is competitive with last spring's closed models rather than this autumn's open leaders. Meituan's table puts it at 70.8 on Terminal-Bench 2.1 and 59.5 on SWE-bench Pro, close to GPT-5.5 (73.8 and 58.6) and behind Claude Opus 4.8 (78.9 and 69.2). GLM-5.3 scores 88.2 on the same Terminal-Bench version, by Z.ai's count. On general reasoning it trails further: 88.9 on GPQA-diamond against 94.3 for Gemini 3.1 Pro.
Its price is the main argument. At $0.30 and $1.20 per million tokens on OpenRouter it costs roughly a quarter of GLM-5.3's $1.40 and $4.40, and the MIT licence carries no revenue clause or naming rule. Meituan also points to support in Claude Code, OpenClaw and Hermes Agent.
The catch: few people can run it themselves. With 1.6 trillion parameters, even the FP8 weights need a multi-node cluster, and Meituan's deployment guide points to SGLang or its own NPU fork. Only one OpenRouter provider serves it, Artificial Analysis lists no index score and measured a 43.53-second wait for the first answer token, and Meituan's newer LongCat-2.5 is API-only, so 2.0 may get little further attention.
Our take: LongCat-2.0 is a cheap second opinion for coding agents and a notable proof that frontier-scale training works off NVIDIA hardware. For a model to build a product on, GLM-5.3 or DeepSeek-V4-Pro have more providers behind them.
Where it earns its place
A startup running a coding agent in Claude Code or OpenClaw
Cut the model bill for repository-level edits
Point the agent at LongCat-2.0 through OpenRouter. At $0.30/$1.20, a session with 400K input and 30K output tokens costs about $0.16 before caching, against about $0.69 on GLM-5.3.
A research group studying training on non-NVIDIA hardware
Inspect a trillion-scale model trained on domestic accelerators
The MIT-licensed weights, LongCat Sparse Attention design and NPU deployment fork are public, so the architecture and serving path can be studied and reproduced on Ascend-class hardware.
Before you commit
- Setup
- Call meituan/longcat-2.0 on OpenRouter, or download BF16, FP8 or INT8 weights and follow the SGLang cookbook for GPUs or Meituan's SGLang-FluentLLM fork for NPUs.
- Integrations
- OpenAI-compatible API through OpenRouter; chat template with tool calls and switchable thinking; Meituan names Claude Code, OpenClaw and Hermes Agent as supported harnesses.
- Data and privacy
- Weights under MIT, so self-hosting keeps data in-house. Hosted traffic goes to AtlasCloud through OpenRouter, which says the model is not available through EU in-region routing.
- Cost at scale
- $0.30/$1.20 per million tokens with cached input at $0.006. Self-hosting a 1.6-trillion-parameter model needs a multi-node GPU or NPU cluster.
- Limits
- 1M-token context and up to 262,144 output tokens on AtlasCloud. Artificial Analysis measured 55 tokens a second and 43.53 seconds to the first answer token.
What’s good
- MIT licence with no revenue or naming conditions
- $0.30/$1.20 per million tokens, $0.006 cached input
- 1M-token context and up to 262,144 output tokens
- Terminal-Bench 2.1 at 70.8 and SWE-bench Pro at 59.5, per Meituan
What’s not
- 1.6 trillion parameters put self-hosting out of reach for most teams
- Only one hosted provider on OpenRouter
- No independent index score yet; slow first token in Artificial Analysis's tests
- Superseded by the API-only LongCat-2.5
Best for
Developers who want a cheap, MIT-licensed large model for coding agents and can live with a single hosting provider.
LongCat-2.0 alternatives
LongCat-2.0: common questions
- Is LongCat-2.0 worth it?
- A trillion-scale open model under a plain MIT licence, cheap to call but already overtaken by its own successor. We rate it 3.8 out of 5.
- How much does LongCat-2.0 cost?
- Open weights under the MIT licence. OpenRouter (AtlasCloud): $0.30 input / $0.006 cached / $1.20 output per 1M tokens.
- What are the downsides of LongCat-2.0?
- 1.6 trillion parameters put self-hosting out of reach for most teams; only one hosted provider on OpenRouter; no independent index score yet; slow first token in Artificial Analysis's tests; superseded by the API-only LongCat-2.5.
- Who is LongCat-2.0 best for?
- Developers who want a cheap, MIT-licensed large model for coding agents and can live with a single hosting provider.
- How much does LongCat-2.0 cost for a team or in production?
- $0.30/$1.20 per million tokens with cached input at $0.006. Self-hosting a 1.6-trillion-parameter model needs a multi-node GPU or NPU cluster.
- Is LongCat-2.0 safe for company data?
- Weights under MIT, so self-hosting keeps data in-house. Hosted traffic goes to AtlasCloud through OpenRouter, which says the model is not available through EU in-region routing.
- What are the best LongCat-2.0 alternatives?
- The closest alternatives we have reviewed are GLM-5.3, DeepSeek-V4-Pro, Kimi K3 and MiniMax-M3.
Answers checked Nov 7, 2026.