China's 1.6-Trillion-Parameter LongCat-2.5 Was Trained Without a Single Nvidia GPU

China's open-weight AI labs keep closing the gap — and this time the hardware story matters as much as the model. Meituan, the Chinese food-delivery and e-commerce giant, has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents, under the permissive MIT licence. It was announced on 2 November 2026 and the weights went public on 9 November. The headline claim: it was trained entirely on Chinese hardware, with no Nvidia GPU involved.

What LongCat-2.5 is

• Mixture-of-experts, 1.6 trillion parameters in total, with roughly 48 billion active per token — so each token is routed through a subset of specialist sub-networks rather than the whole model.
• Built for agents, not chat. It is designed for long, multi-step work done on its own, especially coding: working through a codebase, fixing bugs and checking its own results.
• Multimodal from the start. It converts images into tokens and processes them alongside text, so an agent can look at a screenshot or a chart and act on it — useful for turning a screenshot of a broken page into a fix.
• Long context: up to 1 million tokens at once through the API, with responses up to 128,000 tokens long.

The Chinese-hardware angle

Meituan says it trained the predecessor, LongCat-2.0, in late June 2026 entirely on a cluster of more than 50,000 Chinese AI chips, and that LongCat-2.5's training and fine-tuning also ran on Chinese accelerators. The published self-hosting guide describes a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series — two handling image processing, 64 handling prompt ingestion and 128 handling response generation. There is no guidance for Nvidia GPUs at all. The experts' weights and activations are quantised to 4-bit precision, which Chinese chips handle natively; Meituan reports throughput up 43.3% for prompt processing and 42.77% for response generation against an 8-bit version, with accuracy on par.

That matters because it is the point of the exercise. US export controls were meant to slow Chinese AI by cutting off advanced chips. A frontier-class model trained and served on domestic hardware is exactly the outcome those controls were designed to prevent — and it makes Chinese accelerators more attractive to buyers outside China too.

How it scores — and the fine print

These are Meituan's own numbers, so read them as the vendor's case, not an independent verdict:

• Terminal Bench v2.1 (agents solving command-line tasks): 88.4 — against Claude Opus 5 at 89.1, GPT-5.6 Sol at 88.8 and DeepSeek-V4.1-Flash at 90.6. A near-tie with the previous generation of frontier models.
• Terminal Bench v4.0 (the harder variant): 38.5 — above GPT-5.6 Sol's 37.3 but well below Opus 5's 53.9.
• DeepSWE (fixing bugs in real repositories): 64.6, up from 9.7 for LongCat-2.0 — a huge jump over its predecessor, though DeepSeek-V4.1-Flash leads at 74.2.
• BrowseComp (web search benchmark): 91.9, the highest in Meituan's table — but run in Meituan's own environment.

One detail worth knowing: Meituan benchmarked against GPT-5.6 Sol and Claude Opus 5 even though their successors — GPT-6.1 Sol (29 September) and Claude Opus 5.5 (22 September) — were already out. On the independent Terminal Bench 4.0 leaderboard, both successors score higher than LongCat-2.5. The honest reading is that LongCat-2.5 competes with the previous generation of premium models on selected tasks, at a much lower price — not that it has caught today's frontier.

How to try it

• Chat: try it free in the browser at longcat.ai.
• API: limited-time promotional pricing of $0.30 per million input tokens and $1.20 per million output tokens (cached input $0.006). The API speaks both the OpenAI and Anthropic formats and plugs into agent tools like Claude Code, Codex and Hermes.
• Weights: published on Hugging Face under the MIT licence, which allows commercial use, modification and redistribution. Note the honest distinction: the weights are open, but the training data and code are not fully open — this is open-weight, not open-source.

The catches

• You cannot run it at home. A 1.6-trillion-parameter model needs roughly 3.2 TB for weights alone in 16-bit form. Self-hosting is a data-centre job; the realistic route for almost everyone is the hosted API or the chat interface.
• The training-hardware claims are Meituan's own and have not been independently verified — no training code has been published and no chip count was given for 2.5.
• US scrutiny is real. In October 2026 the US Department of Homeland Security ordered federal civilian agencies to remove DeepSeek technology from covered systems. VentureBeat found no publicly disclosed US federal restriction naming Meituan or LongCat-2.5 as of 9 November — but the rules can change quickly, and anyone serving US government clients should check contract terms before building on it.

Bottom line

LongCat-2.5 is two stories at once. As a tool, it is a genuinely usable open-weight agent model — near-frontier on terminal and coding tasks, multimodal, MIT-licensed, and cheap to try through the API at longcat.ai. As a signal, it is the clearest evidence yet that Chinese labs can now train frontier-class models without Nvidia hardware, which is precisely what export controls were meant to prevent. Try it for screenshot-driven bug fixes and design-to-code work, where its visual understanding earns its keep — and judge it on your own tasks, not on the vendor's benchmark table.

Sources: The Decoder, 15 November 2026, VentureBeat and longcat.ai.

Originally published on Ai Game Changer Pro: https://aigamechangerpro.blogspot.com/2026/11/chinas-16-trillion-parameter-longcat-25.html