LongCat-2.0 — Meituan open-sources a trillion-parameter agentic coder under MIT

Meituan released LongCat-2.0 on June 30 — a large-scale Mixture-of-Experts model, ~1.6 trillion total parameters with ~48B activated per token, 1M context, MIT-licensed and aimed squarely at agentic coding. It was the stealth "owl-alpha" model on OpenRouter, and it is live now via OpenRouter and the longcat.chat platform (billing enabled). Like every trillion-scale model, it changes no local planner pick — but MIT at this size is notable, so it lands as a frontier comparator the way GLM-5.2, Kimi K2.7, and DeepSeek V4-Pro do.

Verdict: A 1.6T MIT MoE built for agentic coding with 1M context — genuinely open-licensed, but trillion-scale keeps it hosted, not a local pick

The take

The facts, verified against Meituan's Hugging Face card + tech blog (longcat.chat/blog/longcat-2.0) and corroborating coverage: 1.6T total / ~48B active MoE, 1M-token context, native tool calling and multi-step reasoning, MIT license. Meituan emphasizes that both training and large-scale deployment run entirely on AI ASIC superpods — a notable infrastructure-independence story. It is accessible today through OpenRouter (the former "owl-alpha") and the longcat.chat API with billing; the open weights are rolling out to GitHub + Hugging Face (the HF weights repo carries the README + MIT LICENSE but the safetensors were still uploading at the time of writing — verify the download before planning around local weights).

Why it matters: a genuinely MIT-licensed trillion-parameter coding-and-agent model is a meaningful open-weight datapoint, and it continues the pattern of Chinese labs (Meituan now alongside Moonshot, DeepSeek, Z.AI, Qwen) pushing the permissively-licensed frontier. Vendor positioning is "superior coding + long-context agent tasks"; treat benchmark claims as vendor numbers until third parties reproduce them.

Our call: no planner-pick change. At 1.6T total it is datacenter / serious-multi-GPU class even at aggressive quantization — not a single-box local model, MIT or not. We track it as a hosted frontier comparator (use it via OpenRouter / longcat.chat today; revisit as a self-hostable option only once the open weights fully land and the community produces a runnable quant). For everyone running locally, coding picks are unchanged: Qwen3-Coder-30B-A3B, North Mini Code, Qwen 3.5 35B-A3B at 24 GB+.

Where this fits

Models: Kimi K2.6 · GLM-5.1 · DeepSeek V4-Pro · Qwen3-Coder-30B-A3B

Hardware: NVIDIA DGX Spark · Dual RTX 5090 · Mac Studio M3 Ultra 96 GB

Sources

Next step

Try this in the planner→