Signal Product Hunt July 2026 — LongCat-2.0, a 1.6T MoE open-source model trained entirely on AI ASICs

Summary

Meituan released LongCat-2.0 on June 30, 2026, a 1.6 trillion-parameter mixture-of-experts model that activates only about 33-56 billion parameters (averaging roughly 48 billion) per token, keeping inference cost well below its headline size. The model was trained from scratch on a 50,000-card cluster of domestic Chinese AI ASICs rather than Nvidia GPUs, spanning more than 35 trillion tokens with no rollbacks or irrecoverable loss spikes — which Meituan calls the industry's first trillion-parameter model to complete full-process training and inference on non-Nvidia hardware. It features a native 1-million-token context window enabled by "LongCat Sparse Attention," which cuts long-context cost from quadratic to linear, plus "zero-computation experts" that skip work for easy tokens. The model was released under a permissive MIT license and had run anonymously as "Owl Alpha" on OpenRouter for two months before being revealed as LongCat-2.0. The same day's Product Hunt digest, covering 16 or more new launches, also featured Ogment AI, which integrates directly into Slack and is activated by tagging "@O"; Katalyst, which automates Salesforce pipeline management through AI agents; and Kuberns, which autonomously deploys and manages cloud infrastructure.