# LongCat-2.0 'beats GPT-5.5' by 0.9 points. That's not the story.

> Meituan's LongCat-2.0, open-sourced in late June 2026, was trained end-to-end on 50,000 Chinese chips — no Nvidia.

*A food-delivery company trained a 1.6-trillion-parameter model with zero Nvidia chips and gave it away. Everyone's reading the wrong number.*

By The InsidersFeed Desk · InsidersFeed
Canonical: https://insidersfeed.com/news/longcat-2-chinese-chips-open-weight

> **Key:** **The take:** the '59.5 beats GPT-5.5's 58.6' line is bait — one vendor-reported benchmark, 0.9 points, unverified. The story you should be filing is that a food-delivery company just trained a trillion-parameter model without a single Nvidia chip. Export controls were built to make that impossible.

Let's dispatch the number everyone quoted. In late June 2026 Meituan open-sourced **LongCat-2.0**, a 1.6-trillion-parameter Mixture-of-Experts model under an MIT licence, and published a benchmark row claiming it edges GPT-5.5. The edge: **59.5 versus 58.6 on SWE-Bench Pro**. That is 0.9 points, on one test, self-reported by the vendor, with no independent confirmation. On any honest reading that is a tie, not a win. If that were the whole story, LongCat-2.0 would be a footnote in the open-weight pile.

## The number that isn't in the benchmark table

Here is the comparison that actually matters, and it has nothing to do with SWE-Bench:

| | LongCat-2.0 | Typical frontier model |
| --- | --- | --- |
| Total params | 1.6T | 1T+ |
| Training silicon | >50,000 domestic ASICs | Nvidia H100/H200 |
| Nvidia GPUs used | Zero | Tens of thousands |
| Licence | MIT | Mostly closed |

Every frontier-scale model you can name was trained on Nvidia. Meituan says LongCat-2.0 was not — trained and served end-to-end on more than 50,000 Chinese-made ASICs. That is the first credible trillion-parameter counter-example to the entire premise of US export controls, which is that denying China Nvidia denies it frontier compute. If Meituan's claim survives scrutiny, that premise just took its first real hit at scale.

> Meituan said LongCat-2.0 was trained entirely on Chinese chips — a cluster of more than 50,000 domestic ASICs with no Nvidia GPUs — and presented it as the first trillion-parameter model built end-to-end on domestic silicon.
> — [VentureBeat](https://venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic-coding-model-thats-been-leading-openrouter-trained-entirely-on-chinese-chips), 2026-06-30

## The 'Owl Alpha' tell

For roughly two months a model called **Owl Alpha** sat at or near the top of OpenRouter's developer usage charts and nobody knew who made it. At launch Meituan admitted Owl Alpha was LongCat-2.0 running incognito. Sit with what that means. Developers picked this model — repeatedly, at volume, with real money — with the brand hidden. No 'trusted lab' halo, no marketing, no country-of-origin bias to lean on. It won on output and price alone. That is the most damning thing in this story for the closed Western labs, because it proves the moat was never the weights; it was the distribution and the brand, and open weights on an OpenRouter endpoint dissolve both.

> **Note:** **The caveat, stated plainly.** The benchmarks are Meituan's own and unverified. At announcement the full weights were listed as 'coming soon', not all posted. The non-Nvidia training claim is Meituan's word until someone reproduces it. Believe the direction of travel; don't yet bank the specific numbers.

## Why this keeps happening

> Meituan released LongCat-2.0 under a permissive MIT licence, allowing free commercial use and self-hosting, and framed the launch as a milestone for domestic compute independence.
> — [Open Source For You](https://www.opensourceforu.com/2026/06/meituan-open-sources-longcat-2-0-under-mit-license/), 2026-06-29

Chinese open-weight models now account for roughly **61% of OpenRouter's top-10 traffic**. That is not an accident and it is not one company — it is DeepSeek, Moonshot, MiniMax and now a food-delivery giant, all shipping capable weights under permissive licences while the US frontier labs meter access and lobby for release controls. Every time Washington tightens the screws on one axis — chips, model exports, release standards — the open-weight side answers on another. LongCat-2.0 answers on the chip axis, the one that was supposed to be airtight. File it correctly: not 'China ties GPT-5.5 on a coding test', but 'China trained a trillion-parameter model without Nvidia and gave it away for free.' The first framing is a leaderboard footnote. The second is a policy problem.

And don't miss the licence detail, because it's the strategic dagger. MIT means anyone — a US startup, a rival lab, a hobbyist — can download LongCat-2.0, run it on their own hardware and build a commercial product on top, with no fee and no phone-home. That's not generosity; it's distribution as strategy. Meituan just put a near-frontier coding model into every developer's reach for free, which does more to erode the US labs' pricing power than any benchmark boast ever could — and it did it while proving the export-control chokepoint can be routed around at the chip level, too.

## Key takeaways

- Meituan open-sourced LongCat-2.0 in late June 2026 — a 1.6-trillion-parameter MoE, MIT-licensed, native 1M-token context, activating ~33-56B parameters per token.
- The take everyone's missing: it was trained and served end-to-end on more than 50,000 domestic Chinese ASICs with no Nvidia GPUs — the first trillion-parameter model built that way.
- The 'beats GPT-5.5' headline is real and hollow: 59.5 vs 58.6 on one vendor-reported benchmark (SWE-Bench Pro), a 0.9-point gap inside the noise, unverified independently.
- LongCat-2.0 spent about two months as 'Owl Alpha', an anonymous model topping OpenRouter's developer charts — it won demand before anyone knew who built it.
- Chinese open-weight models now command roughly 61% of OpenRouter's top-10 traffic; export controls were supposed to prevent exactly this.

## FAQ

### Did LongCat-2.0 really beat GPT-5.5?
Barely, and only on paper. Meituan's own figure is 59.5 on SWE-Bench Pro versus GPT-5.5's 58.6 — a 0.9-point gap on a single benchmark, self-reported and not independently verified. Treat it as a statistical tie, not a win. The benchmark is the least interesting thing about this release.

### Why does the Chinese-chip claim matter so much?
Because US export controls are built on the premise that denying China advanced Nvidia chips caps how large a model it can train. Meituan says LongCat-2.0 — 1.6 trillion parameters — was trained and served on 50,000-plus domestic ASICs with zero Nvidia. If verified, that is the first trillion-parameter crack in the whole export-control theory.

### What was Owl Alpha on OpenRouter?
Owl Alpha was LongCat-2.0 running anonymously. It topped OpenRouter's developer usage charts for about two months before Meituan revealed the origin — meaning developers chose it on merit and price with the brand hidden. That's the strongest evidence in the story that open weights erode the closed labs' real moat: distribution and brand.

### Can I download and use LongCat-2.0 now?
It's MIT-licensed, so commercial use and self-hosting are permitted. But at announcement Meituan listed the full weights as 'coming soon' rather than fully posted, so check the Hugging Face repository for what's actually available before you build on it.

## Sources

- [Meituan open sources LongCat-2.0, the 1.6T near-frontier agentic coding model trained entirely on Chinese chips](https://venturebeat.com/technology/meituan-open-sources-longcat-2-0-the-1-6t-near-frontier-agentic-coding-model-thats-been-leading-openrouter-trained-entirely-on-chinese-chips) — VentureBeat, 2026-06-30
- [meituan-longcat/LongCat-2.0](https://huggingface.co/meituan-longcat/LongCat-2.0) — Hugging Face, 2026-06-29
- [Meituan Open Sources LongCat-2.0 Under MIT License](https://www.opensourceforu.com/2026/06/meituan-open-sources-longcat-2-0-under-mit-license/) — Open Source For You, 2026-06-29