Skip to content
NewsModel release

Meituan open-sources LongCat-2.0, a 1.6T-parameter coding model trained entirely on 50,000 domestic Chinese chips

· by Pondero Newsdesk

The short version

China's food-delivery giant Meituan officially released LongCat-2.0 on June 30, 2026. The 1.6-trillion-parameter MoE model had already reached the top three on OpenRouter by call volume before anyone knew who built it.

Meituan open-sources LongCat-2.0, a 1.6T-parameter coding model trained entirely on 50,000 domestic Chinese chips

A Chinese food-delivery company just revealed it has been running a frontier coding model in stealth for two months, ranking in the top three globally on OpenRouter by call volume, all without a single Nvidia GPU.

What happened

Meituan officially released LongCat-2.0 on June 30, 2026, publishing on GitHub and Hugging Face (weights listed as "coming soon") and simultaneously revealing it as the model behind "Owl Alpha," an anonymous deployment that had spent roughly two months near the top of OpenRouter's coding and agent leaderboards.

The model is a Mixture-of-Experts architecture with 1.6 trillion total parameters. It activates between 33 billion and 56 billion parameters per token at inference (averaging around 48 billion), per the official model page. Native context length is one million tokens, enabled by LongCat Sparse Attention (LSA), a linear-complexity attention variant the team designed to avoid the quadratic scaling wall that causes ordinary transformers to lose coherence past roughly 100,000 tokens.

Training ran on a 50,000-card cluster of domestic Chinese ASICs. Per Meituan's release, no Nvidia GPU was used at any stage of pretraining or large-scale serving. The pretraining corpus exceeds 30 trillion tokens across Chinese, English, multilingual, and code data.

Architecture details worth understanding

Three components distinguish LongCat-2.0 from the previous generation of Chinese open-source models.

First, the zero-computation expert layer: simple tokens route to a zero-compute expert and cost nothing, while complex tokens get more resources. This token-level dynamic activation is what keeps the 1.6T parameter count from becoming prohibitively expensive at inference on agentic workloads where the token mix varies wildly.

Second, the training pipeline called MOPD (Multi-Teacher On-Policy Distillation). Rather than training a generalist from scratch, Meituan started from a supervised fine-tuning checkpoint and branched into three specialist groups (Agent experts for tool use and self-correction, Reasoning experts for multi-hop and STEM tasks, Interaction experts for alignment and hallucination suppression), then distilled all three back into one model. The fusion ran on the domestic cluster.

Third, the inference stack uses large-scale expert parallelism with zero-computation expert integration, designed specifically to keep latency low on trillion-parameter MoE decoding at production scale. Meituan reports steady-state throughput exceeding one trillion tokens per day.

Benchmarks and pricing

Per Meituan's own benchmark page, LongCat-2.0 scored 59.5 pass@1 on SWE-bench Pro, ahead of the 58.6 it cites for GPT-5.5 and the 57.3 for Claude Opus 4.6 (LongCat benchmarks). On Terminal-Bench 2.1 it scored 70.8. These are vendor-reported numbers; independent replication has not been published yet.

API access launched alongside the release. Promotional pricing sits at $0.30 per million input tokens and $1.20 per million output tokens. Standard pricing, once the promotion ends, is $0.75 input and $2.95 output. Both tiers include free cached-context reads, per the pricing comparison page. At promo rates, LongCat-2.0 undercuts Kimi K2.6 (listed at approximately $0.95/M input) by more than a factor of three on input cost alone.

Why it matters for AI tool operators

The Owl Alpha stealth run changes the context here. This was not a paper release or a benchmark drop; the model was in production, accumulating real call volume on OpenRouter for two months before anyone knew the vendor's name. Reaching top-three globally by call volume under an anonymous label is a stronger proof-of-deployment than most open-source releases produce before launch day.

For operators evaluating open-weight coding models, the combination of a 1M-token context window, a competitive SWE-bench Pro score (if independently reproduced), and promo pricing below $0.40/M input on a production-grade API is a meaningful set of options to reassess. The model runs on OpenRouter today; no self-hosting is required to test it at scale.

The chip angle is the longer-term signal. Full training and serving at frontier scale on domestic hardware, without Nvidia, was treated as a stretch goal across the Chinese AI ecosystem as recently as 2024. Meituan's claim that it shipped production traffic at this scale on domestic ASICs is the most concrete evidence yet that the gap is closing.

What to watch next

Two things will determine how much weight this release earns in the broader ecosystem. First, whether Meituan publishes the actual model weights. The Hugging Face listing is live but marked "coming soon," and a genuinely open-weight model invites the independent reproduction runs that vendor benchmarks cannot substitute for.

Second, whether external evaluators replicate the SWE-bench Pro 59.5 score. The benchmark comparisons on longcatai.org are Meituan's own. A confirmed score at that level from a non-Nvidia-trained model would shift the conversation about Chinese frontier labs from a funding and scale story to a genuine capability story.

Sources