A frontier-class coding model, open and Nvidia-free

The list of AI models you can actually run yourself — not rent through an API — just got a serious new entry. On June 30, China's Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter coding model that had been quietly topping usage charts under a codename. It was trained without a single Nvidia GPU. For anyone building software, the interesting part isn't the geopolitics — it's that a near-frontier agentic coder is now weights you can download.

What actually happened

Meituan released LongCat-2.0 with 1.6 trillion parameters and a 1-million-token context window, open-sourced under a permissive MIT license (per VentureBeat). It's a Mixture-of-Experts design: 1.6T total parameters but only ~48 billion active per token, which is what makes a model that large affordable to actually serve. Before the reveal, it ran on OpenRouter under the codename "Owl Alpha" and was handling roughly 10 trillion tokens a month — meaning developers were already leaning on it hard before they knew whose it was.

The hardware story is the headline elsewhere: Meituan says it's the first trillion-parameter model to complete full training and inference on a 50,000-card cluster of domestic Chinese AI chips, no Nvidia involved (per the South China Morning Post). That proves frontier-scale training no longer depends on one company's GPUs — which, over time, is what breaks the compute pricing that flows into every AI bill.

Why it matters for your business

Most AI in your tools today is a rented API call. That's fine until the price changes, the model gets deprecated, your data has to leave your building, or a rate limit throttles you mid-workday. Open weights are the alternative: a model you can host on your own infrastructure, pin to a version that won't shift under you, and run against sensitive code or customer data without shipping it to a third party.

For a small shop, the calculus is honest: you probably won't self-host a 1.6T model on a server in the back office — that still needs real hardware. But you don't have to. Open weights this capable mean you can run it on infrastructure you rent and control, or lean on the growing set of hosts that serve open models at a fraction of frontier-API prices. The direction that matters is options. A year ago, "just use the open model" meant accepting a real quality drop for agentic coding work. LongCat-2.0 is another data point that the gap is closing — and every capable open model is leverage on the price you pay the closed ones.

Sources: VentureBeat, South China Morning Post.