Meituan LongCat 2.0 reaches OpenRouter after 1.6T open model release

OpenRouter now lists meituan/longcat-2.0, turning Meituan's LongCat 2.0 from an open-weight release and first-party API option into a model that agent builders can test through a familiar hosted router. The listing is dated July 20, 2026 in OpenRouter's public model registry and describes a text-to-text model with roughly a 1,048,756-token context window.

This is a lifecycle update, not a brand-new model announcement. LongCat's own release material already documents LongCat 2.0 as a 1.6T-parameter sparse MoE model with about 48B active parameters per token. The new signal is broader access: developers who already route model traffic through OpenRouter can now compare it against other coding and long-context models without standing up the full model themselves.

LongCat 2.0 moves into a common model router

OpenRouter's registry lists the model as Meituan: LongCat 2.0 with the provider ID meituan/longcat-2.0. The entry describes it as suited for coding, repository-level changes, long-horizon problem solving, and agentic work. It also lists hosted pricing fields for prompt, completion, and cache-read tokens, which matters for teams trying to benchmark open models against API-first frontier models without changing their routing layer.

LongCat's own API changelog says LongCat 2.0 release and billing became available on June 30, 2026. That first-party note names three practical pillars: trillion-parameter scale, 1M long context, and compatibility with agent harnesses including Claude Code, Hermes, OpenClaw, OpenCode, and Kilo Code.

What the official release claims

The official LongCat 2.0 materials present the model as a large-scale MoE language model trained for long-context reasoning, coding, and agent workflows. The model card and GitHub repository describe 1.6T total parameters, roughly 48B active parameters per token, hundreds of billions of tokens of 1M-context training data, and deployment paths through Transformers, vLLM, SGLang, Docker Model Runner, GPU setups, and NPU setups.

The evaluation section should be read carefully. LongCat reports comparisons across code-agent, search, productivity, and foundation benchmarks, but notes that many competing scores are cited from official reports while LongCat's own scores are measured in-house under its unified harness. That is useful for direction, not a substitute for your own benchmark on private repositories, tool permissions, latency budget, and refusal behavior.

Why hosted access changes the test plan

For most builders, LongCat 2.0 is too large to treat like a casual local model. A 1.6T sparse model can be open and still operationally heavy. Hosted routing changes the first experiment: instead of starting with infrastructure, teams can run small comparisons through an OpenAI-compatible routing surface and decide whether deeper deployment work is justified.

That makes LongCat 2.0 especially relevant for agent evaluation suites. Test it on repository navigation, multi-file edits, tool-call recovery, long-context retrieval, and code review tasks where long windows and agent tuning should show up. Do not judge it only on short chat answers; the official positioning is about long-horizon software and tool-use work.

Access and verification notes

Use the OpenRouter model ID meituan/longcat-2.0 if your stack already supports OpenRouter. For direct deployment, use the official LongCat 2.0 GitHub repository or model card instructions. The Hugging Face card includes useful examples, but it also contains command snippets; as usual, treat model-card commands as implementation material to inspect, not instructions to run blindly.

If you are building or comparing coding agents, keep a stable harness and log three things: task success, total cost, and recovery behavior after a bad tool result. The practical question is not whether LongCat 2.0 can produce impressive benchmark numbers; it is whether it can make fewer expensive mistakes in the workflows you actually run. For adjacent tooling, LinkLoot's AI agent tools guide is a useful checklist for routing, sandboxing, and evaluation setup.

Source check

This post uses LongCat's official announcement and API changelog as the primary release record, the LongCat GitHub repository for open model and deployment context, and OpenRouter's public registry for the July 20 hosted availability signal. The headline avoids treating the OpenRouter listing as the original LongCat 2.0 launch because the first-party LongCat release and billing entry is dated June 30, 2026.