Meituan LongCat-2.0: 1.6T Open Model Tops Coding

A food-delivery company just shipped one of the most capable open-source coding models on the planet. Meituan — yes, the Chinese super-app you order dinner from — open-sourced LongCat-2.0 in late June, a 1.6-trillion-parameter mixture-of-experts model built for agentic coding, with a native 1-million-token context window and weights on Hugging Face. According to VentureBeat's writeup, it climbed to the top of OpenRouter's usage rankings and was trained entirely on Chinese chips.

That last detail is the story. Plenty of strong open models have come out of China this year. What's different here is the combination: near-frontier coding performance, a permissive open release, and a training run that reportedly never touched an NVIDIA GPU. If it holds up, LongCat-2.0 is the clearest evidence yet that the compute moat isn't what it was.

What Meituan actually released

Here's what's public, per Meituan's official LongCat-2.0 blog post and the Hugging Face model card:

• Scale: 1.6 trillion total parameters in a mixture-of-experts (MoE) architecture. As with all large MoE models, only a fraction of those parameters activate per token — that's how a model this size stays servable. Meituan positions it as "near-frontier," not a claim to beat the very best closed models on everything.
• Context: native 1M-token window, aimed squarely at agentic workloads — repo-scale reasoning, long tool-use chains, multi-file edits — rather than short chat turns.
• Focus: agentic coding. This isn't a general assistant that happens to code; it's tuned for the SWE-bench-style workflow of reading a codebase, planning, editing, and running tools.
• License: open weights released on Hugging Face. That means self-hosting, fine-tuning, and inference through third parties like OpenRouter are all on the table.
• Hardware: Meituan says the model was trained on domestic Chinese accelerators end to end — no NVIDIA in the loop.

The headline isn't the parameter count. It's that a trillion-scale agentic coder got trained without the hardware everyone assumed was mandatory, and then given away.

The benchmark claims — and how to read them

Meituan and early coverage point to strong results on agentic coding benchmarks, including SWE-bench-family evaluations (the research pitch specifically cites SWE-bench Pro). I'd treat the exact numbers with the usual caution: vendor-reported benchmarks are marketing until a neutral third party reproduces them. My read: the more credible signal isn't any single score — it's that LongCat-2.0 rose to the top of OpenRouter's rankings on real developer traffic.

OpenRouter routes production API calls across dozens of models, and its leaderboards reflect what people actually keep paying to use, not what tops a curated eval. When a brand-new open model out of a delivery company starts leading that board, developers are voting with their tokens. That's harder to game than a benchmark table.

Still, "tops coding" needs an asterisk. Leading OpenRouter usage in a given window is partly a novelty-and-price effect: a free-to-cheap open model with frontier-ish quality will spike hard on adoption. Whether it stays at the top after the launch rush is the question that matters, and we don't have that data yet.

Why "trained on Chinese chips" is the whole point

Since the 2022–2025 export-control regime, the working assumption in AI has been that frontier training requires large clusters of high-end NVIDIA GPUs — and that Chinese labs, cut off from the best silicon, would trail by a generation or more. LongCat-2.0 is a direct challenge to that assumption.

We've seen this pressure building all year. DeepSeek made its V4-Pro price cut permanent and kept shipping competitive models. Zhipu's GLM-5.2 topped open coding benchmarks. Now a 1.6T model claims a fully domestic training stack. The trajectory is consistent: Chinese labs are converting a hardware constraint into an efficiency-and-openness strategy.

The honest take: "trained on Chinese chips" is a real milestone, but it's also a talking point, and the specifics matter more than the slogan. Which accelerators? What cluster size? What throughput and cost versus an equivalent NVIDIA run? Meituan's materials emphasize the achievement more than the engineering economics, and until independent researchers dig into the training details, we can't say whether this is compute parity or a heroic one-off that cost far more than a comparable Western run. The claim is important either way — it just isn't the same as proof that the chip gap has closed.

How it stacks up against the open-model field

LongCat-2.0 lands in an unusually crowded open coding market. A rough map of the moment:

Two things stand out. First, three of the four strongest recent open coding releases are Chinese — that's not a coincidence, it's a strategy. Second, LongCat-2.0 is the heaviest of the group by a wide margin. The others compete on efficiency and price; Meituan competes on raw capability and context length. If you're self-hosting, GLM-5.2 or a small model like North Mini Code is far cheaper to run. LongCat-2.0 is the one you reach for API-hosted, when you want frontier-adjacent agentic coding without paying frontier closed-model prices.

What this means for developers right now

Practically, you don't need to care where a model was trained to use it. What you get with LongCat-2.0 is another top-tier option on OpenRouter and a downloadable set of weights if you have the infrastructure to serve a 1.6T MoE (which is not most teams — this is a multi-GPU, serious-hardware model to self-host, so realistically you'll consume it via an API provider).

For anyone building coding agents, the useful questions are:

• Does the 1M context hold quality across the window? Big context numbers are easy to advertise and hard to deliver. Long-context models often degrade well before the stated limit. This needs independent long-context evals before you architect an agent around it.
• How does it behave inside a real agent loop? SWE-bench scores don't fully capture tool-calling reliability, instruction-following under long horizons, or how gracefully it recovers from a failed step. That's where agentic coders actually live or die.
• What's the latency and cost at 1.6T? A model this large is not free to run even when the weights are. Per-token pricing through hosted providers is the number that decides whether it fits your budget, and that will vary by provider.

If you're already routing through tools like Cursor, GitHub Copilot, or a custom OpenRouter setup, the near-term impact is simple: you get one more strong model to A/B against your current default. Open weights also mean the community can fine-tune domain-specific variants, which is where open releases tend to compound over the following months.

The geopolitics under the hood

Zoom out and LongCat-2.0 fits a pattern I keep coming back to on this site. Export controls were designed to slow Chinese frontier AI. The observable response has been the opposite of surrender: Chinese labs are shipping capable models and open-sourcing them, which does two things at once. It builds global developer mindshare on Chinese-origin models, and it undercuts the pricing power of closed Western labs by giving away something close to their product.

I think that's the real strategic move here. Whether or not LongCat-2.0 is genuinely "near-frontier" by every measure, open-sourcing a trillion-scale agentic coder trained on domestic silicon is a statement: the compute chokepoint is being routed around, and the route runs straight through the open-source ecosystem Western labs increasingly avoid. Meanwhile, closed labs are moving the other direction — restricted previews, government-first access, tighter gates. The contrast is getting hard to miss.

What we don't know yet

Being honest about the gaps: we don't have independent, reproduced benchmarks confirming the coding claims. We don't have detailed training-hardware specs — which specific Chinese accelerators, how many, at what cost and efficiency versus an NVIDIA baseline. We don't know how durable the OpenRouter lead is past the launch spike. And "open weights" covers a range of licenses; the exact commercial terms are worth reading on the Hugging Face card before you build a business on top of it.

None of that diminishes the release. A 1.6T open agentic coder, trained without NVIDIA and leading real developer usage on day one, is a genuinely notable event regardless of how the fine print shakes out. My bottom line: test it against your current coding model on your own tasks before you believe any leaderboard — but do test it. When a delivery company ships a trillion-parameter open model that developers immediately start using in production, the interesting part isn't the cat pun. It's that the barrier to entry for frontier-class coding AI just got a lot lower, and it got lower in Hangzhou, not Silicon Valley.