Product Hunt · AIJul 7, 2026, 14:27 UTC+8

Meituan releases LongCat-2.0: a 1.6T MoE model, MIT-licensed, trained on custom AI ASICs

LongCat-2.0

Meituan launched LongCat-2.0 on Product Hunt: an MIT-licensed 1.6T-parameter MoE model with ~48B active parameters and 1M context window. It uses LongCat Sparse Attention and is post-trained for coding and agentic workflows. The model was trained entirely on Meituan's own AI ASIC superpods, not NVIDIA GPUs. It integrates with Claude Code, OpenClaw, and Hermes. The post doesn't disclose benchmark scores, API pricing, or throughput — I'd hold off on performance claims until numbers land.

Why it matters: Meituan LongCat-2.0 is a 1.6T-param MoE model, MIT-licensed, trained entirely on in-house AI chips with a 1M-token context window and post-training focused on code and production deployment. Flagship domestic model release with a non-NVIDIA training story — HKR all hit. No ben...

Read the original ↗Export Markdown

Take

Two numbers jump out: 1.6 trillion total parameters with only 48 billion active, and a 1-million-token context window. The wilder part is the training stack — Meituan ran this entirely on its own AI ASIC clusters, zero NVIDIA GPUs. It's MIT-licensed and already wired into Claude Code, OpenClaw, and Hermes, so the target is clearly coding and agent workflows.

I'd hold off on getting excited until we see numbers. The Product Hunt page has no benchmarks, no API pricing, no throughput specs. 48B active parameters is roughly DeepSeek-V3 territory — the real question is performance per dollar and inference speed. Meituan hasn't disclosed anything about its chip specs either, so we can't gauge training efficiency from the outside.

What's concrete: Meituan is betting hard on both its own models and its own silicon, and it chose the most permissive license possible. Whether LongCat-2.0 actually competes depends entirely on the benchmarks they haven't published yet.

Source: Product Hunt · AI · producthunt.com
Published: Jul 7, 2026, 14:27 UTC+8 · Jul 7, 2026