Meituan, the Chinese food-delivery giant, open-sourced one of the largest coding models ever built. On June 30, 2026 it released LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts model tuned for agentic coding, under a permissive MIT license, and the headline is not the size. It is that Meituan trained the whole thing on a 50,000-card cluster of domestically produced Chinese chips, with no Nvidia H100s and no AMD MI300X anywhere in the pipeline.

  • 1.6 trillion parameters, MoE. LongCat-2.0 activates only 33 to 56 billion parameters per token, so it runs far cheaper than a dense model of the same nominal size.
  • Trained on Chinese silicon. A 50,000-card domestic cluster handled the full run, which Meituan calls a first of its scale for Chinese hardware.
  • It was already number one. The model topped OpenRouter developer usage for two months in stealth under the codename "Owl Alpha" before Meituan unmasked it.
  • Open weights, cheap API. Weights are on Hugging Face under meituan-longcat; promotional API pricing starts at $0.30 per million input tokens with free context-cache hits.
How a Mixture-of-Experts model activates a fraction of its parameters LongCat-2.0 holds 1.6 trillion parameters total but routes each token to only 33 to 56 billion active parameters. A 1.6T model that only fires 3% at a time Total parameters 1,600B Active per token: 33-56B dormant experts stay cold A router picks a handful of experts for each token, so compute cost tracks the orange sliver, not the full bar. Context window stretches to 1,000,000 tokens. Benchmarks (vendor-reported): SWE-bench Pro 59.5 | Terminal-Bench 70.8 genztech.blog
Fig 1 Mixture-of-Experts keeps the model enormous on disk but cheap at inference: only the orange sliver of parameters runs for any given token.

What exactly did Meituan release?

A frontier-adjacent open-weights coding model and the tooling to run it. LongCat-2.0 uses a Mixture-of-Experts architecture, which means it does not push every token through all 1.6 trillion parameters. Instead a routing network selects a small set of specialist sub-networks per token, activating roughly 33 to 56 billion parameters at a time. That is the trick behind the whole release: you get the capacity of a colossal model with the running cost of a mid-sized one. The context window reaches 1 million tokens, enough to load a large repository and reason across it in a single pass, which is exactly what agentic coding workloads demand. The weights, tokenizer, and inference code all live on Hugging Face and GitHub under the meituan-longcat organization.

RelatedClaude Fable 5 Returns and Retakes the Coding Crown

Why does training on Chinese chips matter so much?

Because it is the clearest public evidence yet that US export controls are not the hard ceiling Washington intended. Since late 2022 the United States has restricted China's access to top-tier AI accelerators, betting that compute scarcity would slow Chinese frontier models. Meituan trained a 1.6-trillion-parameter model on a 50,000-card domestic cluster with none of the restricted parts. The performance claims are Meituan's own, and independent verification of both the benchmark scores and the hardware story would strengthen confidence, but the direction is unmistakable: domestic Chinese silicon has crossed the threshold where training-scale runs are viable, not just inference.

TraitLongCat-2.0Typical dense flagship
Total params1.6T (MoE)Hundreds of B (dense)
Active per token33-56BAll of them
LicenseMIT (open weights)Usually closed
Training hardwareDomestic Chinese chipsNvidia/AMD
Context1,000,000 tokens128K-1M

How good is it, really?

Good enough to top a real usage chart before anyone knew who made it. For two months LongCat-2.0 ran on OpenRouter as the anonymous "Owl Alpha" model and quietly led developer usage, which is a market signal that no benchmark table can fake. On the numbers Meituan reports 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, and it claims parity with Google's Gemini 3.1 Pro from February. Those figures sit below the current closed-source leaders, but for an MIT-licensed model you can download and self-host, the price-to-capability ratio is the story. Promotional API access starts at $0.30 per million input tokens with free context-cache hits, a fraction of what closed frontier models charge.

Why is Meituan building AI at all?

Survival economics. Meituan's core delivery business faces brutal domestic competition and compressing margins, so the company publicly committed billions to AI and domestic chip capability as a way to move up the value chain. LongCat-2.0 is the third rung on a fast-climbing ladder: LongCat-Flash, a 560-billion-parameter model, shipped in September 2025; LongCat-Next, a multimodal variant, followed in March 2026; and LongCat-2.0 nearly tripled the parameter count in under a year. A delivery app becoming a serious open-model lab in three releases is its own kind of statement about how cheap and fast this game has become.

RelatedClaude Sonnet 5 Nearly Matches Opus at Half the Price

What to watch · 2026
  • Independent evals. The benchmark and hardware claims are vendor-reported. Watch vals.ai and the official SWE-bench leaderboard for third-party confirmation.
  • Adoption after the reveal. "Owl Alpha" led on mystery. Watch whether usage holds now that the Meituan name and Chinese-chip provenance are public.
  • Export-control response. A training-scale run on domestic silicon undercuts the compute-embargo thesis. Watch for policy reaction.
  • The next rung. Meituan tripled scale in a year. Watch how quickly LongCat-3 or a multimodal 2.0 lands.

Our take

LongCat-2.0 matters less as a model and more as a proof of concept: a Chinese company trained a trillion-scale system on domestic hardware and shipped it under the most permissive license there is. The benchmark scores are respectable rather than record-breaking, and the smart move is to treat Meituan's self-reported numbers as a starting hypothesis until independent evaluators weigh in. But the two months it spent silently topping OpenRouter as Owl Alpha is the part that should get attention, because real developers picked it on merit without knowing its lineage. If you build agentic coding tools, an MIT-licensed 1.6T model with a million-token context and $0.30 input pricing is worth a serious evaluation. And if you were betting that export controls would keep Chinese labs a generation behind, this release is the data point arguing otherwise.

Primary sources

Original analysis by GenZTech. Figures current as of July 2026. Source: huggingface.co/meituan-longcat