A food-delivery company just shipped one of the most geopolitically loaded AI models of the year. On June 30, 2026, Meituan released LongCat-2.0, a 1.6-trillion-parameter open-weight coding model under the MIT license. The parameter count is not the story. The training silicon is.
Meituan says the model was trained end-to-end on roughly 50,000 domestic Chinese AI ASIC chips with zero Nvidia hardware involved. That claim is confirmed across trade-press coverage from Reuters, SCMP, and VentureBeat.
If it holds, LongCat-2.0 is the largest AI model trained entirely outside the Nvidia supply chain, and it becomes a real data point in the export-control debate rather than another press cycle.
The detail that makes practitioners pay attention: the preview build was quietly topping real usage on OpenRouter before the stable launch.
TL;DR
LongCat-2.0 is a 1.6T-parameter Mixture-of-Experts coding model, MIT-licensed, with about 48B active parameters per token and a 1M-token context window. Meituan trained it on a ~50,000-chip domestic cluster with no Nvidia.
The architecture and infrastructure facts are well-sourced. The benchmark numbers are not, because there is no technical report yet. Adopt it for self-hosting under strict data governance, and verify everything else on your own codebase.
Key takeaways
- LongCat-2.0 ships under the MIT license with full downloadable weights, confirmed in the meituan-longcat/LongCat-2.0 GitHub repository and the Hugging Face model card.
- It is a sparse MoE: 1.6T total parameters, only ~48B active per token, so each forward pass touches roughly 3% of the network.
- Training ran on ~50,000 domestic Chinese AI chips. The vendor is unconfirmed, with Huawei Ascend the leading rumor.
- Every benchmark number is vendor-reported. No arXiv paper exists as of June 30, 2026, which breaks Meituan's prior publication pattern.
- The "leading OpenRouter usage" claim is real but contested, with token-volume rankings and category rankings telling different stories.
What is LongCat-2.0, in one paragraph?
LongCat-2.0 is Meituan's open-weight agentic coding model: a 1.6-trillion-parameter Mixture-of-Experts system that activates roughly 48 billion parameters per token, carries a 1-million-token context window, and is released under the permissive MIT license. It is built for autonomous software engineering, meaning multi-file edits, terminal operations, and test-driven loops rather than single-line autocomplete.
That puts it in the same competitive lane as Anthropic's Claude Code, OpenAI's Codex, and Cursor, with the difference that you can download and self-host the weights.
How does the architecture actually work?
The sparsity is the interesting part. With 1.6T total parameters but only ~48B active per token, LongCat-2.0 gets the representational capacity of a giant model while paying inference cost closer to a 48B dense model.
The active figure is worth pinning down precisely, because some early write-ups floated a 33B-to-56B range. First-party sources triangulate to about 48B.
That 1M-token context window matters for the agentic use case. An autonomous coding agent needs to hold an entire repository, a long test log, and several turns of its own edits in working memory. A short context forces constant re-retrieval and breaks multi-step plans.
Meituan's lineage here is real. The team published LongCat-Flash (arXiv:2509.01322) and LongCat-Flash-Thinking (arXiv:2509.18883) in September 2025, then LongCat-Video (arXiv:2510.22200) in October and LongCat-Image (arXiv:2512.07584) in December. Going from LongCat-Flash to a 1.6T model in roughly nine months is a steep scaling curve.
The chip sovereignty claim, and what's verified
This is the part the policy world cares about. US export controls have restricted Nvidia's H100 and successor chips from Chinese entities since 2022. Chinese firms responded by building out domestic silicon, mainly Huawei's Ascend line plus custom ASICs from Cambricon and Biren.
What is solidly confirmed about LongCat-2.0's training:
- It used domestic Chinese AI ASIC chips, no Nvidia.
- The cluster was approximately 50,000 chips.
- Training was end-to-end: pre-training, post-training, and inference all on domestic hardware.
What is not confirmed:
- The exact chip vendor. Huawei Ascend 910C is the common guess, but Meituan has not named it in any verified source.
- The effective compute in FLOPS versus an Nvidia cluster of similar headcount. Chip count is not compute.
- Whether training on this hardware costs quality versus an equivalent Nvidia run.
The US-China Economic and Security Review Commission's March 2026 report, titled "How China's Open AI Strategy Reinforces Its Industrial Dominance," frames China's open-weight push as deliberate industrial policy: give domestic enterprises frontier-grade models without US-controlled infrastructure. LongCat-2.0 is a clean test case for that thesis.
Near-frontier performance suggests the controls are not biting. A clear gap suggests they are.
Is LongCat-2.0 really "leading OpenRouter usage"?
Partly, and the conflicting data is worth understanding before you cite it. The preview model, codenamed "Owl Alpha," launched on OpenRouter for testing on April 24, 2026. One widely repeated figure, reported by KuCoin's news desk, puts it at about 10.1 trillion monthly tokens and first place globally.
A separate read of the data shows Owl Alpha at 6th place with 4.1% share in OpenRouter's Programming category specifically.
Both can be true. Total platform token volume and category-specific ranking measure different things. A cheap, fast model can rack up enormous token counts on bulk workloads without dominating any quality leaderboard.
Independent testing on Benchable.ai placed Owl Alpha in the 16th-to-19th percentile for speed and general knowledge, and community reviews on Linux.do called it "average, not a reasoning model."
So the durable statement is this: LongCat-2.0's preview drove real, heavy production traffic, which is a genuine adoption signal that most open-weight launches never produce. Check the live OpenRouter rankings yourself before repeating any specific placement.
Why you should distrust the benchmark numbers
Meituan reports strong scores. There is no technical report to back them, and the benchmark ground beneath them is shifting.
The claimed numbers, all vendor-reported and unreplicated:
| Benchmark | Claimed LongCat-2.0 | Comparison |
|---|---|---|
| SWE-bench Pro | ~59.5 | vs GPT-5.5 ~58.6 |
| Terminal-Bench 2.1 | ~70.8 | vs competitors |
| SWE-bench Multilingual | ~77.3 | , |
| FORTE | ~73.2 | , |
Two problems. First, no arXiv paper exists for LongCat-2.0 as of June 30, 2026, which breaks Meituan's own pattern of publishing one for every prior LongCat release. Without it, the architecture, training recipe, and evaluation method cannot be audited.
Second, the SWE-bench family lost credibility this year. OpenAI deprecated SWE-bench Verified on February 23, 2026 after finding that 59.4% of audited hard tasks had broken test suites.
SWE-bench Pro is Scale AI's private set, so you cannot inspect it either. A ~59.5 that beats GPT-5.5 by under a point, on an unverifiable private benchmark, with no paper, is not a number to make decisions on.
How it compares to the 2026 Chinese open-weight field
LongCat-2.0 did not arrive alone. It landed in the busiest year yet for Chinese open weights.
| Model | Org | Key specs | Released |
|---|---|---|---|
| LongCat-2.0 | Meituan | 1.6T total, ~48B active, MIT, 1M ctx | Jun 2026 |
| MiniMax M3 | MiniMax | ~1T+, multimodal | 2026 |
| Moonshot Kimi K2.6 | Moonshot | ~200B+, 1M ctx | 2026 |
| Zhipu GLM-5.2 | Zhipu | ~100B+ | 2026 |
| DeepSeek-V3 | DeepSeek | ~236B active, MoE | Dec 2025 |
Stanford HAI's DigiChina project reads this diversity as a distinct strategic bet against the Western closed-frontier approach. LongCat-2.0's specific contribution is the no-Nvidia training story plus the genuinely permissive license.
What this means for you
Start with the one scenario where LongCat-2.0 is clearly worth a pilot: your code cannot leave your infrastructure. If data governance, GDPR, or a security review forbids sending source to an external API, the MIT-licensed weights let you self-host the whole thing. Claude Code and Codex structurally cannot offer that.
For everyone else, here is the decision table.
| Your situation | Call | Why |
|---|---|---|
| Code can't leave your infra | Strong candidate | MIT weights, full self-hosting |
| Want near-frontier coding quality | Test it yourself | Benchmarks unverified, no paper |
| Need enterprise SLA and support | Stay on incumbents | No support tier confirmed |
| Need reproducible benchmarks | Not yet | No technical report to audit |
| Agentic workflows, low stakes | Pilot it | 1M context and MoE fit the job |
Budget the real cost before committing to self-hosting. At BF16, ~48B active parameters need roughly 96GB just for weights, so you are looking at multiple accelerators per instance plus a serving stack like vLLM or SGLang, autoscaling, and monitoring.
Against that, weigh the unknown API pricing on longcat.chat and the compliance cost of shipping code off-site. The community is already moving, with Unsloth adding LongCat-2.0 support for fine-tuning.
The honest summary: LongCat-2.0 is a credible engineering achievement with a thin evidence trail. The infrastructure and licensing are real and immediately useful. The performance claims are marketing until a paper or your own evals say otherwise.
Treat it as a serious self-hosting option and a weak source of benchmark bragging rights, and you will read it correctly.
Sources
- Meituan LongCat-2.0 GitHub repository (meituan-longcat/LongCat-2.0)
- LongCat-2.0 model weights and model card on Hugging Face (meituan-longcat)
- LongCat official site and API platform (longcat.chat)
- LongCat-Flash technical report, arXiv:2509.01322
- LongCat-Flash-Thinking technical report, arXiv:2509.18883
- LongCat-Video technical report, arXiv:2510.22200
- LongCat-Image technical report, arXiv:2512.07584
- US-China Economic and Security Review Commission, "How China's Open AI Strategy Reinforces Its Industrial Dominance," March 2026
- OpenAI statement deprecating SWE-bench Verified, February 23, 2026
- Stanford HAI DigiChina project brief on China's open-weight ecosystem
- Reuters, SCMP, and VentureBeat reporting on the no-Nvidia training claim
- Benchable.ai independent testing of the LongCat-2.0 preview ("Owl Alpha")