Slide 1
Meituan is a food-delivery company. It just open-sourced LongCat-2.0, a one-point-six-trillion-parameter model trained on AI accelerators instead of Nvidia GPUs. The weights are already public, under MIT.
The open weights behind "Owl Alpha"
For two months, this model quietly ran near the top of OpenRouter's agent leaderboards under the codename Owl Alpha. Now the name and the weights are public: a one-point-six-trillion-parameter mixture-of-experts, released under MIT.
Trained end-to-end on AI accelerators, not GPUs
The real story isn't a benchmark. It's how it was built. Pretraining ran across more than thirty-five trillion tokens, entirely on purpose-built AI accelerators, with no rollbacks and no irrecoverable loss-spike restarts. A frontier-scale model, trained end to end off the GPU. Only about forty-eight billion of its parameters fire per token, which is what makes that feasible.
SWE-bench Pro: 59.5, past GPT-5.5, short of Opus
On SWE-bench Pro, an agentic coding benchmark, LongCat scores fifty-nine point five. That's ahead of Gemini three point one Pro and GPT five point five, though still behind Claude Opus four point seven and four point eight. And every rival number there is the vendor's own reported score, not a test Meituan re-ran head-to-head.
One sparse-attention scheme keeps 48B-active cheap at long context
Two design choices keep it cheap at long context. A sparse-attention scheme that learns which tokens are worth attending to, and a hundred and thirty-five billion extra embedding parameters that add capacity outside the expert mix. It's trained on data with context up to a million tokens long.
What the leaderboard win leaves out
Three honest caveats. This isn't state of the art. It trails Claude Opus on every code-agent benchmark, and on GPQA-diamond, pure knowledge, it scores eighty-eight point nine, below every frontier model here. The cross-vendor comparison isn't controlled, and the hardware story is Meituan's own claim, with no chip vendor named and no outside audit.
Grab the weights
The weights and inference code are on Hugging Face now, under the MIT license, so you can run it or build on it today. The full tech report is on Meituan's blog.
Slide 8
The headline here isn't a new leaderboard king. It's that a frontier-scale open model was trained end to end without a single Nvidia GPU. If that holds up, the map of what you can build these models on just got a lot wider.
