Meituan releases LongCat-2.5, a 1.6-trillion-parameter agent model trained on Chinese hardware

Meituan has released the model weights for LongCat-2.5, a massive multimodal agent model trained entirely on domestic Chinese hardware.

2 min read

Meituan releases LongCat-2.5, a 1.6-trillion-parameter agent model trained on Chinese hardware

Chinese food delivery company Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built specifically for autonomous AI agents. The model was trained entirely on domestic Chinese hardware and approaches the performance of leading Western systems on several coding benchmarks.

Key facts

Total parameters
1.6 trillion
Active parameters
48 billion
Context window
1 million tokens
Max output length
128,000 tokens
License
MIT
Input price
€0.27 per million tokens
Terminal Bench v2.1 Performance Comparison
Score
LongCat-2.5: 88.4
Claude Opus 5: 89.1
GPT-5.6 Sol: 88.8
Deepseek-V4.1-Flash: 90.6
Bar charts comparing LongCat-2.5 performance with Claude Opus and GPT models.
Benchmark performance comparisons between LongCat-2.5 and leading Western AI models across multiple coding and general agent tasks., Source ↗

Architecture and Capabilities

LongCat-2.5 is a mixture-of-experts model — an architecture where the system routes each token to specialized sub-networks to save computing power. Out of 1.6 trillion total parameters, roughly 48 billion are active for each token processed. The model handles text, images, and video, processes up to one million tokens via its API, and generates responses up to 128,000 tokens long. Users can also toggle reasoning capabilities on and off.

Instead of bolting on a separate image encoder, LongCat-2.5 converts images directly into tokens that are processed alongside text. For training, the developers used real software tasks from GitHub pull requests, blocked internet access in training environments, and distilled capabilities from several specialized teacher models.

Diagram showing the LongCat-2.5 scaling agentic RL loop and vision tokenization steps.
Overview of the LongCat-2.5 reinforcement learning and tokenization pipeline., Source ↗

Hardware and Deployment

Meituan quantizes the expert weights and activations to 4-bit precision, a format that Chinese chips can process natively. A sample self-hosted deployment runs on a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. Two chips handle image processing, 64 handle prompt ingestion, and 128 handle response generation. The self-hosting guide contains no configuration guidance for Nvidia graphics processing units.

The model weights are available under the MIT license. Meituan offers the model through an API that supports both OpenAI and Anthropic formats, alongside a chat interface at longcat.ai.

Benchmark Performance

On Terminal Bench v2.1, a test measuring command-line task resolution, LongCat-2.5 scores 88.4 points, trailing models like Claude Opus 5 and GPT-5.6 Sol. On DeepSWE, which evaluates bug fixing in real code repositories, the model scores 64.6 points. However, it achieves 91.9 points on BrowseComp, a web search benchmark.

Source: Original announcement

Worth a read?

Report

Comments