Meituan releases open weights for LongCat-2.5, a 1.6T-parameter MoE agent model trained on Chinese hardware, under MIT license
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Meituan published the weights of LongCat-2.5 (1.6T total parameters, ~48B active per token, 1M-token context, 128K max output, text/image/video input) under the MIT license, one week after its official unveiling. A separate 4-bit quantized version is provided with a self-hosting guide for a 194-chip Huawei Ascend Atlas A2 cluster, with no Nvidia guidance. Meituan's API (OpenAI- and Anthropic-compatible) is offered at a promotional price of about €0.27 per million input tokens and €1.07 per million output tokens, and a chat interface is available at longcat.ai.
Entities: Meituan, LongCat-2.5, LongCat-2.0, LongCat-Next, LongCat-Image, Huawei Ascend Atlas A2
0 primary
What happened
Meituan has published the weights of LongCat-2.5 under the MIT licence, one week after unveiling it. It is a mixture-of-experts model with 1.6 trillion total parameters, about 48 billion active per token, a 1M-token context window, 128K maximum output, and text, image and video input. Meituan also supplies a 4-bit quantised version and a self-hosting guide for a 194-chip Huawei Ascend Atlas A2 cluster, with no Nvidia guidance. Its API, compatible with OpenAI and Anthropic formats, is at a promotional price of about €0.27 per million input tokens and €1.07 per million output tokens.
Why it matters
The main point is the combination of a permissive licence and a documented non-Nvidia deployment path. Enterprises and developers, particularly in China or in places with chip export limits, get a usable template for running a very large model on Huawei hardware. The licence lets anyone self-host or fine-tune the model commercially, and the low API price puts pressure on rival providers. The practical impact is limited, though. A 194-chip cluster is far beyond most buyers, and nobody outside Meituan has yet shown that the model matches top US models on real work.
What is noise
The headline claim that the model "never touched an Nvidia GPU" and approaches top US models on agent benchmarks rests on Meituan's own figures, relayed by secondary coverage. Benchmark results from the maker are not independent evidence, and benchmark scores often fail to carry over to real agent tasks. The API price is promotional, so it says little about long-run costs, and the "frontier-level" label is unproven until others reproduce it.
Watch next
- 01Independent evaluations and community leaderboard results on agent and coding tasks, to test whether Meituan's benchmark claims against Claude Opus 5 and GPT-5.6 Sol hold up
- 02Third-party reports of anyone reproducing the 194-chip Ascend deployment, with real throughput, latency and cost per token, and whether the 4-bit version loses quality
- 03Pricing after the promotional period ends, plus usage and ranking on OpenRouter, and any move by Huawei or other Chinese labs to ship comparable Ascend-first open releases
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680