Meituan open-sourced 1.6T-parameter LongCat-2.0 trained on 50,000 domestic ASICs, claiming China's first full-process domestic-chip trillion-parameter model.
Meituan open-sourced LongCat-2.0, a 1.6 trillion-parameter LLM trained entirely on domestic chips. The model claims to be China's first trillion-parameter AI fully pre-trained and inferred on a 50,000-card ASIC cluster.
Key facts
- 1.6 trillion parameters in LongCat-2.0.
- 1 million-token context window.
- 50,000-card domestic ASIC cluster used for training.
- DeepSeek V4-pro also has 1.6 trillion parameters.
- Open-sourced on Tuesday by Meituan.
Food delivery giant Meituan on Tuesday open-sourced LongCat-2.0, a large language model boasting 1.6 trillion parameters and a 1 million-token context window According to SCMP. The Beijing-based company claimed this is the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster built with AI ASIC superpods.
Beyond Inference
While DeepSeek's V4-pro (1.6 trillion parameters, launched April 2026) relied on home-grown chips only for inference, Meituan says LongCat-2.0 used domestic hardware for both pre-training and inference. Pre-training is far more computationally intensive — it involves digesting massive datasets to learn basic patterns. This marks a significant step for China's push to move domestic chips beyond inference workloads.
The Hardware Question
Meituan did not disclose the specific ASIC vendor or chip performance metrics. The claim of a 50,000-card cluster raises questions about interconnect efficiency and training stability at scale on non-Nvidia hardware. DeepSeek's V4-pro, by contrast, used domestic chips only for inference — a less demanding task — while likely relying on Nvidia or other foreign GPUs for pre-training, though DeepSeek has not confirmed that.
Open-Source and Context
LongCat-2.0 is open-sourced, following Meituan's earlier LongCat-1.0 release. The 1 million-token context window matches frontier models like DeepSeek V4 (which achieved 500K context with FlashMemory optimization in June 2026) and positions LongCat for long-document and enterprise RAG use cases. Meituan has not published benchmark results on standard evaluations like MMLU, HumanEval, or SWE-Bench.
What to watch
Watch for benchmark results from Meituan on standard evaluations like MMLU, HumanEval, and SWE-Bench. Also track whether DeepSeek responds with a fully domestic-chip pre-training claim for its next model, potentially V5.
Source: scmp.com
Originally published on gentic.news
Top comments (0)