Meituan released LongCat-2.5 on November 2, positioning it as a natively multimodal model built for long-horizon agent tasks. The headline number is 1.6 trillion total parameters. The more interesting number is that the entire training run happened on domestic Chinese accelerators.

This is not a first attempt

LongCat-2.0 in June was the first Meituan model to complete full training and inference of a trillion-parameter model on domestic silicon. Version 2.5 builds on that base: stronger native visual understanding, stronger agentic coding, and better long-context training and inference efficiency on the same hardware class.

"Domestic chips, domestic model" sounds like a slogan. In engineering terms it means the architecture is adapted to the memory and interconnect characteristics of those cards, the training framework rewrites communication paths, and inference is partitioned around the reality of small per-card memory.

What was actually engineered

Per Meituan's technical blog, mid-training used sequences up to 1M tokens, with trillions of tokens of cumulative training. Long sequences across many cards hit two walls immediately: memory capacity and inter-card communication.

Meituan's answer was co-optimization with the chip vendor across memory usage, interconnect and the balance of compute versus transfer. To work around small per-card memory, task partitioning moved into the supernode so that parameter and activation traffic stays on the high-speed fabric inside it.

The measured result: on long-context supervised fine-tuning, single-card training throughput improved 86 percent over a distributed data parallel baseline. Put differently, one card now does the work of roughly 1.86 cards in the same wall-clock time.

Where it lands on capability

Published evaluations place LongCat-2.5 between Claude Opus 4.8 and Claude Opus 5, with some benchmarks above Opus 5.

Treat that with the usual caution. Benchmark scores swing several points with prompting, sampling and tool-calling implementations. The honest test is your own task set.

Two capability areas matter for applications. Native multimodality converts visual input into tokens the model handles uniformly, which supports tasks like interactive web pages and 3D product pages. Agentic coding uses large-scale long-horizon reinforcement learning to improve end-to-end engineering work.

How to try it

Consumers can use it at longcat.ai. Developers can call the API, which launched alongside the model.

For teams evaluating a switch, test in this order: long-context tasks first, since that is the stated differentiator; then multimodal page reconstruction; then agentic coding completion rate. Clear all three before discussing migration.

The real test is the ecosystem

Training a trillion-parameter model on domestic silicon moved from "possible" to "efficient" in about a year of engineering. The harder problem is what comes next: operator coverage, tooling maturity, and developer habits. Those decide whether a model gets used or just gets benchmarked.