Meituan LongCat-2.0
1.6T open MoE coding agent trained wholly on Chinese chips
Overview
LongCat-2.0 is Meituan's 1.6-trillion-parameter mixture-of-experts LLM (~48B active per token) tuned for agentic coding, with a native 1M-token context. Its headline is engineering as much as size: the first frontier-scale model trained and served end-to-end on domestic Chinese chips, not Nvidia.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 13 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with Meituan LongCat-2.0
- 1.6T-parameter MoE, ~48B active per token
- Native 1M-token context (LongCat Sparse Attention)
- MIT-licensed open weights (staged release)
- Tuned for agentic coding / IDE-copilot workloads
- GPU and NPU deployment support
- Available via third-party APIs (e.g. OpenRouter)
ASEAN Perspective
Meituan LongCat-2.0 in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
The genuinely notable thing here isn't a benchmark — it's the supply chain. Meituan says LongCat-2.0 was trained and run entirely on roughly 50,000 domestic accelerators (analysts point at Huawei Ascend or Cambricon) with no Nvidia or AMD in the loop, which if it holds up is the first trillion-parameter model to pull that off. On capability it's strong but not a clean frontier win: Meituan's own numbers put it ahead of a Claude Opus reference on IFEval (90 vs 86) and IMO-AnswerBench, but behind on SWE-bench Pro (59.5 vs 69.2) — and these are self-reported, with no independent Epoch AI run yet. It earned real credibility running anonymously as 'Owl Alpha' on OpenRouter for two months, topping agent-workspace charts by call volume. Caveat worth knowing: the MIT license was announced but full weights were still 'coming soon' at reveal, so verify availability before you build on it.
What people say
Independent reviews are thin because the model is new and, at least at announcement, still API-only — but the pre-reveal track record does the talking.
For roughly two months before Meituan's July 6, 2026 announcement, LongCat-2.0 ran on OpenRouter under the alias 'Owl Alpha,' quietly racking up about 10.1 trillion tokens a month and climbing agent-focused leaderboards: reportedly first on the Hermes Agent workspace, second on Claude Code, and third across OpenClaw deployments by call volume. That's usage-based validation from developers who didn't know whose model they were paying for, which carries more weight than a vendor benchmark table.
The architecture is the other talking point. It's a 1.6-trillion-parameter MoE that activates an average of ~48B parameters per token (dynamically 33B-56B depending on query complexity), paired with 'LongCat Sparse Attention' to deliver a native 1M-token window at linear rather than quadratic cost. Meituan claims full 1M access across all layers rather than a sliding-window approximation.
On benchmarks, treat Meituan's figures as self-reported. Their tables show LongCat-2.0 beating a Claude Opus reference on IFEval (90.0 vs 86.0) and IMO-AnswerBench (81.8 vs 75.3) but trailing on SWE-bench Pro (59.5 vs 69.2). No Epoch AI or other third-party evaluation had been published at the time of writing, so the 'near-frontier' framing in the trade press (VentureBeat, Dealroom, AI Weekly) rests largely on Meituan's own numbers plus the OpenRouter usage signal.
The hardware claim is what most coverage leads with: training on a cluster of around 50,000 domestic compute cards with no Nvidia or AMD GPUs, using the Huawei Collective Communication Library for training stability. Meituan didn't name the chip vendor; analysts nominate Huawei Ascend or Cambricon.
Two practical caveats for anyone planning to deploy: despite the MIT-license announcement, both the GitHub and Hugging Face pages listed weights as 'coming soon' at reveal, meaning it was effectively hosted-API-only; and because it's optimized for agentic coding, it's a specialist pick rather than a general assistant. Check weight availability and run your own eval before committing.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including Meituan LongCat-2.0's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Meituan LongCat-2.0 unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Meituan LongCat-2.0 directly →
Spotted something out of date? Suggest an update →
Alternatives to Meituan LongCat-2.0
More in LLMs & Chat
