Artificial Intelligence

LongCat-2.5 Open Source Release: Meituan's 1.6T Multimodal Agent Model with 1M Context

By 美团技术团队 · Nov 10, 2026

Read original on juejin.cn ↗ Google Translate ↗ Alt translation

A 1.6T-parameter MoE model with 1M context is now freely modifiable for commercial use, and its training and quantization methods offer a working path for running long-sequence workloads on constrained hardware. Western teams evaluating open multimodal agent models can test LongCat-2.5 against their own code and office pipelines without licensing friction.

Summary

Meituan released LongCat-2.5 as an MIT-licensed open-source model for long-horizon agent tasks. It has 1.6 trillion total parameters, activates roughly 48 billion per token, and supports a one-million-token context. The model improves end-to-end delivery in complex codebase development, full project builds, visual coding, and professional office scenarios.

The training pipeline was redesigned to get there: longer real-world trajectories, partial rollout asynchronous RL with dynamic oversampling, and gradient correction on stale checkpoints. Visual input is turned into multi-level discrete tokens using residual vector quantization, so vision and text share one backbone. W4A4 quantization keeps inference cost down, and a series of FSDP and communication optimizations lets the model train and fine-tune on domestic AI chips with less than 80GB of single-card memory.

The license allows commercial inference and retraining without extra permission. Meituan also published the model on Hugging Face, a technical blog, and an API platform.

Takeaways

— MIT license allows commercial inference deployment and retraining without additional authorization.

— Total parameters are 1.6 trillion, with about 48 billion active per token, and context length is one million tokens.

— Improvements target complex codebase development, full project builds, visual coding, and professional office scenarios.

— Training uses partial rollout asynchronous RL, dynamic oversampling, and gradient correction on stale trajectories.

— Visual tokens are produced via residual vector quantization on native-resolution features, sharing the same backbone parameters as text.

— W4A4 quantization quantizes expert weights and activations to 4-bit with native INT4 matrix multiplication, keeping downstream performance comparable.

— FSDP dual sharding and hybrid vocabulary parallelism enable long-sequence training on chips with less than 80GB per-card memory; single-card throughput is up 86% vs DDP baseline in long-sequence SFT tasks.

— Model, technical blog, and API platform are available at the provided Hugging Face, longcat.ai, and longcat.chat URLs.

Conclusions

Meituan's choice of MIT license removes the biggest adoption barrier for a model of this scale, so its value will be determined by what the community builds, not by the parameter count.

The visual tokenization approach makes vision a first-class citizen in the same parameter space as text, which simplifies multi-modal agent loops by avoiding a separate encoder.

The focus on sub-80GB memory constraints reveals that the practical bottleneck for trillion-parameter MoE deployment is memory bandwidth, and the engineering choices are tailored specifically to that reality.

Concepts & terms

Partial rollout

An asynchronous reinforcement learning training scheme that saves unfinished long trajectories and resumes them later, allowing long-horizon tasks to be trained without restarting from scratch.

Residual vector quantization

A technique that encodes visual features into multi-level discrete tokens by quantizing residuals layer by layer, letting the backbone model process vision and text in a shared semantic space.

W4A4 quantization

A weight-activation quantization scheme that compresses both expert weights and activations to 4-bit and runs computations on native INT4 matrix multiplication units, reducing inference cost while preserving accuracy.

Tree Attention

An attention mechanism that processes multiple samples in the same group within a single forward and backward pass, improving training efficiency for reinforcement learning with a shared tree structure.

Source: juejin.cn ↗ Google Translate ↗ Backup ↗