Meituan Releases LongCat-2.5 Weights Under MIT: the 1.6T-Parameter, ~48B-Active Native Multimodal MoE Lands on Hugging Face With a W4A4 Quantized Version

Background and Pain Points

When Meituan launched LongCat-2.5 on November 2, it was API-only, and the weights sections on Hugging Face and GitHub said "coming soon." A week later, the official Meituan LongCat account announced on X that the weights are out under the MIT license. For teams that want to self-host, deploy privately or keep training, this adds another trillion-parameter-scale open agent model that can be used commercially. Discussion in the Chinese community has focused on two points: a 1.6T model is rare among open releases, and the MIT license puts almost no limits on commercial use.

Architecture Highlights

LongCat-2.5 is a native multimodal MoE with 1.6 trillion total parameters and about 48 billion active per token. According to the model card, it builds on LongCat-Next by turning native-resolution visual features into discrete tokens aligned with text in a shared semantic space. Training combines long-horizon agentic reinforcement learning, multi-teacher distillation and long-sequence optimizations, all on domestic AI accelerators. W4A4 quantization brings both expert weights and activations down to INT4; Meituan says that versus a W8A8 baseline it raises prefill throughput by 43.3% and decode throughput by 42.77%, with accuracy comparable to the unquantized model on the evaluated benchmarks.

Benchmarks

These numbers come from the model card. LongCat scores are measured in-house and are indicative only; competitor scores marked with an asterisk are cited from public reports. Code agents: Terminal Bench v2.1 88.4 (Claude Opus 5 89.1, DeepSeek-V4.1-Flash 90.6), Terminal Bench v4.0 38.5 (Claude Opus 5 53.9), and DeepSWE v1.1 64.6 (DeepSeek-V4.1-Flash 74.2). General agents: BrowseComp 91.9, AutomationBench 51.8, MCP Atlas 83.1 and HLE with tools 63.1. Overall it is close to or ahead of closed flagships on browsing and automation tasks, but still behind Claude Opus 5 and DeepSeek-V4.1-Flash on pure coding. There are no independent third-party tests yet.

Getting Started

The weights are huge. The BF16 main repository is about 3.56TB and the W4A4 version about 1.34TB, and the model card only offers an NPU deployment guide based on SGLang-FluentLLM, with no GPU instructions, so running it on consumer graphics cards is not realistic. Most developers can still try it through the official API. Teams with NPU clusters that need private deployment can start by sizing hardware costs with the W4A4 version. When calling the model, choose thinking mode (keeping all reasoning content is recommended for better results) or non-thinking mode for better token efficiency; it supports an OpenAI-style API with tool calling.