Meituan Open-Sources a 1.6-Trillion-Parameter AI Agent Model Built Without Nvidia, Scoring Within a Point of Claude Opus 5 on Terminal Bench at a Fraction of the Price

Meituan, the Chinese food delivery giant, is now shipping one of the largest openly licensed AI models in the world, and it never needed an Nvidia GPU to do it. The company’s LongCat team has put the full weights of LongCat-2.5, a 1.6-trillion-parameter model built for AI agents that grind through long, multi-step coding and office tasks, on Hugging Face under the permissive MIT license. As The Decoder detailed on November 15, the model was trained on Chinese hardware, and Meituan’s own self-hosting guide targets Huawei accelerators rather than Nvidia chips.

The release matters well beyond China. A model that sits close to the best American systems on agent benchmarks, can be downloaded and used commercially for free, and is sold through an API at a fraction of Western prices puts fresh pressure on how businesses budget for AI and on how much US chip export controls can slow Chinese competitors.

What Meituan Released

According to the LongCat-2.5 model card, the system is a native multimodal mixture-of-experts model with 1.6 trillion total parameters, of which about 48 billion are active for any given token. That design keeps running costs far below what the headline size suggests, because only a slice of the network does work on each step.

• Inputs and outputs: The Decoder reports that through the API the model accepts up to one million tokens of context, writes responses of up to 128,000 tokens, understands text, images and video, and lets users switch step-by-step reasoning on or off.
• Licensing: Both the main checkpoint and a compressed version carry the MIT license, which allows commercial use, modification and redistribution.
• Size on disk: The Marginal Token counted about 3.56 TB of weight files for the main checkpoint and about 1.34 TB for the compressed 4-bit package.
• Timeline: The Marginal Token dates the launch to November 2, and Meituan’s technical team blog announced the open-source release on November 10.

Rather than bolting a separate image encoder onto a language model, LongCat-2.5 turns pictures into discrete tokens that it processes alongside text, an approach Meituan describes in its launch blog as a way to strengthen visual coding, design and 3D work.

How It Stacks Up Against OpenAI and Anthropic

Meituan’s benchmark table on Hugging Face puts LongCat-2.5 within striking distance of top US models on several agent tests, though the company notes that scores it measured itself are “indicative only.”

• Terminal Bench v2.1, which tests how well an agent completes command-line tasks: LongCat-2.5 scored 88.4, against 89.1 for Claude Opus 5, 88.8 for GPT-5.6-Sol and 90.6 for DeepSeek-V4.1-Flash.
• Terminal Bench v4.0, a harder variant: 38.5, slightly ahead of GPT-5.6-Sol’s 37.3 but well short of Claude Opus 5 at 53.9.
• DeepSWE, a bug-fixing test on real code repositories: 64.6, up from just 9.7 for LongCat-2.0, but behind DeepSeek-V4.1-Flash at 74.2 and GPT-5.6-Sol at 72.7.
• BrowseComp, a web research benchmark: 91.9, the top score in Meituan’s table, although The Decoder notes Meituan ran that test in its own environment.

The Decoder also points out that newer models such as Claude Opus 5.5 and GPT-6.1 Sol are missing from the comparison, so the fair reading is that LongCat-2.5 matches where the US frontier stood a few months ago rather than where it stands today. That lag is shrinking, and it now comes with open weights.

Built on Huawei, Not Nvidia

The hardware story is the part investors and policymakers will watch most closely. Meituan compresses the model’s expert weights and activations to 4-bit precision, a format The Decoder says Chinese chips handle natively. In Meituan’s load tests, that version raised throughput by 43.3% for reading prompts and 42.77% for generating answers compared with an 8-bit version, with accuracy roughly unchanged.

The published self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series: two for image processing, 64 for ingesting prompts and 128 for generating responses. There is no guidance for Nvidia GPUs. Meituan has said it trained the predecessor, LongCat-2.0, on a cluster of more than 50,000 Chinese AI chips, and that model quietly ran on OpenRouter under the name “Owl Alpha” before its official debut, according to The Decoder.

The model card adds that infrastructure work on these domestic accelerators delivered 86% higher per-device throughput than a standard baseline in long-sequence training, which suggests Chinese labs are learning to squeeze more out of hardware that is less powerful chip-for-chip than Nvidia’s best.

Why It Matters for Business

Price pressure is real. The Decoder reports a limited-time promotional API price of roughly €0.27 per million input tokens and €1.07 per million output tokens. The Marginal Token found the hosting service ZenMux listing promotional rates of $0.30 and $1.18, with regular rates of $0.75 and $2.95. Those numbers are a small fraction of what premium US models typically charge, and they give procurement teams leverage when negotiating enterprise AI contracts.

Switching costs are low. Meituan’s API speaks both the OpenAI and Anthropic formats and plugs into agent tools such as Claude Code and Codex, The Decoder says. A developer can test it by changing an endpoint rather than rewriting software, which makes it easy for cost-conscious companies to route routine agent work to cheaper models.

Open weights cut both ways. Companies can run LongCat-2.5 on their own infrastructure and keep sensitive data in-house, but self-hosting a model whose compressed weights alone take up 1.34 TB is a serious engineering project. Firms in regulated industries will also need to weigh data-governance and supply-chain policies before deploying a Chinese-built model, even an openly inspectable one.

Export controls are leaking. Meituan’s training approach, which relied on real GitHub pull requests, sandboxes with internet access blocked and Git history removed, and a separate model checking answers for shortcuts, shows Chinese labs competing on technique, not just raw compute. A frontier-class open model that runs on Huawei silicon makes Chinese chips more attractive to buyers outside China, too, which is the opposite of what US export restrictions were designed to achieve.

The Bottom Line

LongCat-2.5 is not the best AI model in the world, but it does not need to be. By pairing near-frontier agent performance with an MIT license, low API prices and a hardware stack that skips Nvidia entirely, Meituan has made the economics of AI agents more competitive. For businesses, that means more choice and lower costs. For US chipmakers and model labs, it is one more sign that the gap with Chinese rivals is measured in months, not years.

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous