A 1.6T Agent Model Trained Without Nvidia? Inside LongCat-2.5

According to The Decoder, Chinese food delivery giant Meituan has released the weights for LongCat-2.5, a massive 1.6-trillion-parameter Mixture-of-Experts model built for complex agent workflows. What sets this release apart is the complete absence of Nvidia hardware during its training and deployment.
How does the hardware setup actually work?
The self-hosting guide from Meituan details a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. Out of these, 2 handle image processing, 64 process prompt ingestion, and 128 manage response generation.
To make this non-Nvidia infrastructure viable, Meituan quantizes the experts' weights and activations to 4-bit precision, which local Chinese chips process natively. In load tests, this quantization boosts throughput by 43.3 percent for prompts and 42.77 percent for responses without sacrificing accuracy compared to an 8-bit version.
What are the performance tradeoffs?
LongCat-2.5 activates roughly 48 billion parameters per token. On Terminal Bench v2.1, it scores 88.4 points, placing it just behind Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8. However, on bug-fixing benchmarks like DeepSWE, it lags significantly behind Western counterparts, scoring 64.6 points while top models exceed 72 points.
How to test or access the model
Try the chat interface directly at longcat.ai.
Download the open-weights under the MIT license for local or custom cluster self-hosting.
Integrate the API using standard OpenAI or Anthropic formats directly into agent frameworks like Claude Code or Codex.
Is it worth changing your infrastructure stack?
For teams locked into the Nvidia ecosystem, migrating to Huawei Ascend chips just for this model makes little practical sense due to hardware availability and support barriers outside China. However, the model's API compatibility and low promotional pricing — roughly €0.27 per million input tokens — make it a viable alternative for evaluating massive agent architectures without massive cloud costs.
Sources
Frequently asked questions
- Was any Nvidia hardware used to train LongCat-2.5?
- No. According to Meituan's documentation reported by The Decoder, the model was trained entirely on Chinese hardware and runs on Huawei Ascend accelerators.
- What is the token capacity of LongCat-2.5?
- Through its API, the model handles up to one million input tokens at once and generates responses up to 128,000 tokens long.
- What license does LongCat-2.5 use?
- The model weights are released under the MIT license, permitting commercial use, modification, and redistribution.
AI Briefing in 3 Minutes
Get what actually matters in AI & tech delivered free to your inbox. Zero spam, 1-click unsubscribe.
Comments
0 commentsDeixe seu comentário
Be the first to comment.
Continue Reading

Runway's Praxis-1: Can Video AI Actually Control Real Robots?
Runway libera Praxis-1 para controle robótico

Stop Scouting Creators Manually: Nectar’s New Agent Does It For You
Nectar launches AI agent for creator discovery

United Airlines Just Blocked AI Agents: What Now?
United Airlines bans autonomous AI booking agents