Meituan LongCat-2.0 open-weights model hits 1M-token context with 1.6T sparse parameters
Meituan's LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts model with 48 billion active parameters, is now open-source with million-token context and agent-focused architecture.
A new open-weight model built for agentic programming just landed with a context window that handles a million tokens and parameter efficiency that keeps inference costs in check.
LongCat-2.0 is Meituan's 1.6-trillion-parameter mixture-of-experts model that activates roughly 48 billion parameters per forward pass. The architecture splits workload across specialized expert groups—one for agentic tasks, one for reasoning, and one for interaction—using MOPD (mixture-of-expert-per-domain). Sparse attention keeps the million-token context tractable, and ScMoE (sparse conditional mixture-of-experts) gates which experts fire for each token. The model was trained from scratch on more than 35 trillion tokens and supports deployment on both GPU and NPU hardware.
Benchmark results show 59.5 on SWE-bench Pro, 70.8 on Terminal-Bench 2.1, and 77.3 on SWE-bench Multilingual. Those scores place it in the upper tier for code-agent tasks, especially on the multilingual variant where many models still struggle. Weights and model card are available on ModelScope, with inference scripts included for practitioners running it locally or integrating it into agent frameworks that need long-context memory.
