Key Takeaways:
Meituan's LongCat-2.0 proves frontier AI can be built without Nvidia — and it's already the most-used coding model on OpenRouter.
Meituan Inc. open-sourced LongCat-2.0, a 1.6-trillion-parameter AI model trained entirely on 50,000 domestic Chinese ASICs, bypassing Nvidia Corp.'s restricted GPUs to deliver near-frontier coding performance at a fraction of the cost.
"This closes the inference cost gap and proves domestic silicon can scale to frontier levels," Wang Xing, Meituan's chief executive officer, said in a statement.
The Mixture-of-Experts model activates an average of 48 billion parameters per token — ranging from 33 billion to 56 billion depending on query complexity — and supports a 1-million-token context window through a novel LongCat Sparse Attention mechanism. On SWE-bench Pro, LongCat-2.0 scored 59.5, narrowly surpassing OpenAI's GPT-5.5 at 58.6, while posting 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual. The company released BF16, FP8 and INT8 precision versions alongside inference code optimized for domestic chips.
The open-source release under an MIT license threatens to accelerate a structural shift in AI infrastructure spending. If Chinese conglomerates can consistently train trillion-parameter models on homegrown ASICs rather than Nvidia's general-purpose GPUs, the data center GPU market faces a viable alternative ecosystem — one Washington's export controls were designed to prevent.
The Owl Alpha Trail
LongCat-2.0 operated anonymously on OpenRouter as "Owl Alpha" for two months before Meituan claimed the architecture. During that period, the model processed approximately 10.1 trillion monthly tokens — 559 billion per day — representing a 242% month-over-month volume surge that propelled it into the platform's global top three. It secured the top ranking on the Hermes Agent workspace, second place on Claude Code deployments, and third place across international OpenClaw environments.
The model's post-training layer, called Multi-Teacher Optimization via Mixture of Specialized Experts (MOPD), segregates optimization into three independent clusters: Agent Experts for tool invocation and self-correcting loops, Reasoning Experts for multi-hop logic and mathematics, and Interaction Experts for human alignment and safety guardrails. A dynamic gate-routing mechanism fuses these behaviors at runtime, allowing the model to coordinate deep reasoning, stable tool execution and safe interaction simultaneously.
Pricing That Undercuts the Market
Meituan's commercial framework introduces aggressive pricing designed to capture developer mindshare. Standard pay-as-you-go API rates are $0.75 per million input tokens and $2.95 per million output tokens — already below OpenAI's GPT-5.6 Luna at $1.00/$6.00 and Anthropic's Claude Opus 4.8 at $5.00/$25.00. A limited-time promotion slashes those rates to $0.30 and $1.20 respectively, matching Xiaomi's MiMo-V2.5 Flash at the low end.
The structural advantage lies in context cache economics. In massive agentic environments where a coding assistant repeatedly reads and modifies the same multi-million-token code repository, only cache-miss inputs and final token generations consume the quota. This architecture fundamentally alters the cost calculus for large-scale autonomous software development.
The Domestic Chip Imperative
The training milestone carries geopolitical weight. Washington's export controls have restricted Nvidia's most advanced GPUs from reaching China, pushing companies like Meituan, Huawei Technologies Co. and Alibaba Group Holding Ltd. to develop domestic alternatives. Bernstein estimated in 2025 that Nvidia held roughly 40% of China's AI chip market, roughly matched by Huawei, with Nvidia's share expected to fall 8% this year.
Meituan said the LongCat-2.0 cluster was built around large-scale ASIC superpods using Huawei's Collective Communication Library to manage chip-to-chip coordination — mirroring how Nvidia's NCCL coordinates its own GPU clusters. The company acknowledged that memory limits were the primary bottleneck during pre-training, as domestic accelerators carry less memory per device than Nvidia's banned H800 chip.
For investors, the implications cut both ways. Nvidia shares trade at roughly 35 times forward earnings, and a credible domestic alternative in China could compress the premium the market assigns to its data center GPU monopoly. Conversely, Meituan — which started as a Groupon-style deals site in 2010 and now serves 770 million annual transacting users — is positioning itself as an AI infrastructure provider rather than just a food delivery super app. The company's pivot into AI began with the $281 million acquisition of startup Light Year Beyond in 2023 and accelerated with the LongCat-Flash release in late 2025.
This article is for informational purposes only and does not constitute investment advice.
