Beijing tech giant Meituan has made a true breakthrough by unveiling LongCat-2.0 — a large language model with 1.6 trillion parameters. The key feature that has already caught the industry's attention is that its entire training cycle was completed exclusively on domestic equipment, without using a single Nvidia chip. This is China's first model of such scale to undergo a full pre-training cycle on local accelerators.

While many Chinese companies, including DeepSeek, used local chips only for the inference stage (running the finished model), Meituan went further. LongCat-2.0 was trained on a cluster of 50,000 domestic ASIC accelerators. For comparison, DeepSeek V4-pro, a flagship model with a similar context window of 1 million tokens, relied on Western chips during the pre-training stage.

Technological Sovereignty and "Hardware" Challenges

The company confirmed that Huawei's communication library (HCCL) was used to ensure training stability and scalability. This indicates that the infrastructure is likely based on Huawei Atlas-950 superclusters. Renowned analyst TP Huang has already stated that this success removes all doubts about the ability of these clusters to train large language models.

However, the path to this achievement was fraught with difficulties. According to Meituan itself, the main problem was memory. The memory capacity per device of domestic accelerators is significantly smaller than that of the Nvidia H800 chip, which is banned for export to China. Overcoming this limitation required enormous engineering efforts to create a stable and scalable infrastructure.

Benchmarks and Market Positioning

LongCat-2.0 has already shown impressive results in programming and agent tasks, surpassing the earlier Google Gemini 3.1 Pro model in Terminal-Bench 2.1 and SWE-Bench Pro tests. Nevertheless, Meituan honestly admits that their creation still lags behind global flagships — OpenAI GPT-5.5 and Anthropic Claude 4.8 Opus. Additionally, the model has not yet been evaluated on leading independent platforms and the latest comprehensive tests.

Cryptalist Expert Opinion: The launch of LongCat-2.0 is not just a technological success for one company. It is a demonstration that China can create competitive AI solutions even under strict sanctions and without access to advanced Nvidia GPUs. For the cryptocurrency and blockchain sector, this is a signal of a possible acceleration in the development of decentralized computing networks and AI agents based on alternative hardware. However, for now, the "software ecosystem" around Chinese chips remains weaker than the mature CUDA platform, which hinders mass adoption.