---
title: "Meituan open-sources 1.6-trillion-parameter agent model trained without Nvidia"
url: https://www.parallelquant.com/posts/meituan-open-sources-1-6-trillion-parameter-agent-model-trained-without--82010e
source_name: "The Decoder"
source_url: https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/
published: 2026-11-15T10:06:06.000Z
topics: ["open source", "llms", "chips"]
publisher: "Parallel Quant"
---

# Meituan open-sources 1.6-trillion-parameter agent model trained without Nvidia

*2026-11-15 · Source: [The Decoder](https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/)*

Meituan released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and trained entirely on Chinese hardware. It scores 88.4 on Terminal Bench, just behind Claude Opus 5. Meituan's sample deployment runs on 194 Huawei Ascend chips.

**Why it matters:** A near-frontier agentic model trained with no Nvidia GPUs is the strongest evidence yet that US export controls are no longer a hard ceiling on Chinese AI. It lands amid reports of Blackwell fingerprints in Chinese models and widening chip-smuggling probes — if Huawei's Ascend stack can do the job, demand for diverted Nvidia chips may shrink. Open weights at this scale also put pressure on Western open-model releases.

**Topics:** open source, llms, chips

---
Read the original: https://the-decoder.com/meituan-releases-a-1-6-trillion-parameter-agent-model-that-never-touched-an-nvidia-gpu/
Canonical: https://www.parallelquant.com/posts/meituan-open-sources-1-6-trillion-parameter-agent-model-trained-without--82010e
Published by Parallel Quant — https://www.parallelquant.com
