On June 30, 2026, Chinese technology company Meituan released LongCat-2.0 under the MIT open-source license. That alone would be worth noting — 1.6 trillion parameters, competitive with near-frontier models, free to use commercially. What makes it significant is the hardware behind it.

LongCat-2.0 was trained entirely on a cluster of over 50,000 domestically built Chinese ASICs. Not a single Nvidia GPU in the training stack.

Export controls on advanced AI chips were supposed to slow China’s AI development. LongCat-2.0 is the clearest demonstration yet that those controls have not stopped it.

What LongCat-2.0 Actually Is

LongCat-2.0 uses a Mixture-of-Experts (MoE) architecture. The 1.6 trillion total parameters sounds enormous — and it is — but the model only activates around 48 billion parameters per token. That keeps inference costs much closer to what you would pay for a mid-sized dense model while keeping the overall knowledge and capability of a much larger one.

The context window sits at one million tokens. For practical terms, that means the model can hold an entire large codebase in its working memory in a single request, which is directly relevant for the agentic coding use cases it was built to handle.

Before the open-source release, LongCat-2.0 was leading performance rankings on OpenRouter for coding tasks — measured against deployments from OpenAI, Anthropic, and Google. It was not a regional curiosity running inside Meituan’s internal tools. It was the top model on a platform where developers actively compare and route traffic based on actual output quality.

The Chip Story Is the Real Story

US export controls restricted Nvidia’s highest-end H100 and A100 chips from reaching China in late 2022 and 2023. The assumption embedded in those controls was that without access to the best training hardware, Chinese labs would fall behind in developing frontier-capable models.

That assumption has not held.

What happened instead: Chinese companies accelerated domestic chip development. Meituan’s training cluster used ASICs purpose-built for large-scale AI workloads. These are not general-purpose chips pretending to do AI work — they were designed for this. The 50,000-chip cluster represents a sustained, expensive, and now clearly successful bet on domestic AI compute.

The geopolitical implication is direct: the export controls did not prevent China from building a frontier model. They may have slowed the pace, but LongCat-2.0 shows that near-frontier capability is achievable without US hardware. Future export policy debates need to account for that demonstrated outcome, not the assumption that was made in 2022.

What Open-Sourcing Under MIT Means

The MIT license is about as permissive as software licenses get. You can use LongCat-2.0 in a commercial product, fine-tune it, modify it, build a business on top of it, and distribute derivative works — all without paying Meituan anything, all without asking permission.

The weights are coming. At time of writing, the GitHub and Hugging Face repositories list the weights as “coming soon,” but the model architecture and technical documentation are already public. When the weights land, any organisation with sufficient inference infrastructure will be able to run a near-frontier coding model at cost.

For businesses evaluating self-hosted AI options, this is meaningful. The open-source ecosystem has produced capable models before, but LongCat-2.0 has demonstrated production-level performance on real developer workloads, not just academic benchmarks.

What This Means for Business

For most businesses, LongCat-2.0 is not a reason to immediately change AI providers or rebuild your workflows. But it reinforces several important trends worth tracking.

The open-source ceiling keeps rising. A year ago, the conventional wisdom was that open-source models lagged closed models by 6-12 months of capability. That gap has narrowed substantially. LongCat-2.0, Llama’s continued development, and Alibaba’s Qwen series all show that organisations with the right infrastructure can run near-frontier models without a vendor subscription.

AI capability is no longer the moat. When frontier-adjacent models are available under MIT license, the competitive advantage in AI shifts away from model access and toward data, context, and integration. The business that wins is not the one that picked the strongest model — it is the one that fed the model the right information and built the right workflows around it.

Vendor diversity is a real option now. Businesses concerned about dependency on OpenAI or Anthropic have more credible fallback options than they did 12 months ago. That does not mean you should abandon a working platform. It does mean that the risk calculation around AI vendor lock-in has changed.

Coding and knowledge work automation are accelerating fastest. LongCat-2.0 was specifically built and evaluated for agentic coding tasks. The trend of models optimised for tool use, multi-step reasoning, and autonomous work completion is accelerating. Businesses with software development or data operations functions should be actively exploring what that means for their team structure over the next 18-24 months.

Want the practical version of this? The free Working With Claude field guide covers the full Claude ecosystem, Claude Code, and how to roll it out across a real business. Download it here.