Meituan open-sources 1.6T coding model built on Chinese chips

Meituan released LongCat-2.0 under an MIT license, a 1.6-trillion-parameter Mixture-of-Experts model built for agentic coding. It scores 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, and a preview version already ranked top three by call volume on OpenRouter, per VentureBeat.

The number that matters isn't the benchmark score. It's the training cluster: roughly 50,000 domestic chips, zero Nvidia A100s or H100s, zero AMD MI300X. Meituan built a frontier-adjacent coding model without any of the hardware that export controls were designed to block.

Efficient by necessity

LongCat-2.0 uses what VentureBeat describes as a Zero-Compute Experts router, activating only about 33 to 56 billion parameters per token out of the full 1.6 trillion. That's a MoE design choice, but it's also what you'd build if your chip supply was constrained and every FLOP had to count. The model also supports a 1M-token context window, the kind of spec that makes it usable for whole-repo agentic work instead of isolated code completion.

For builders comparing open coding models, this changes the calculus. A model that's MIT-licensed and trained outside the Nvidia and AMD supply chain removes a dependency a lot of teams didn't know they had. It's one more open option backed by real benchmark numbers, not just weights and a README.

A related report covered frontier access getting gated to government-approved partners elsewhere. Meituan went the opposite direction, publishing weights anyone can run without a gate.

What's not in VentureBeat's reporting: pricing for hosted inference, or how the 1M-token context holds up on real repositories rather than benchmarks. The architecture answers the compute question. It doesn't answer the reliability one.