Meituan Launches LongCat-2.5, a 1.6 Trillion Parameter Native Multimodal Model

dev.to Machine Learning2026-11-03 18:150 views

Meituan LongCat-2.5 — 1.6 trillion parameter native multimodal model

On November 2, 2026, Meituan released LongCat-2.5, a native multimodal model built for long-horizon agent tasks, trained entirely on domestic Chinese compute — the largest training run yet completed on local accelerators.

Scale and Architecture

LongCat-2.5 has 1.6 trillion total parameters, with about 48 billion activated per token. Meituan says the model extends the LongCat line by strengthening native visual understanding and agentic coding, representing visual signals as multi-level discrete tokens so the backbone processes vision and text in a single unified token stream. The team optimized long-context training and W4A4 inference on domestic hardware, and used large-scale long-horizon reinforcement learning to improve end-to-end engineering on complex codebases.

Benchmark Results

In public evaluations the model scores between Anthropic's Claude Opus 4.8 and Opus 5, and beats Opus 5 on some agent benchmarks. It posts 51.8 on AutomationBench versus Opus 5's 50.3, 91.9 on BrowseComp for first place, and 55.4 on 3DCodeBench for text-to-3D generation. Meituan frames the result less as a parameter record and more as proof that a trillion-scale training loop can be closed entirely on Chinese accelerators.

Where It Ships

Rather than chasing a general token factory, Meituan embeds LongCat-2.5 into its own business workflows — merchant analytics and agentic coding — keeping the model close to real local-life-service decisions where fault tolerance is low and the payoff is operational.