For weeks, developers on OpenRouter noticed something weird. A model called "Owl Alpha" kept showing up in leaderboards, punching above its weight on coding benchmarks. Nobody knew who made it. Turns out it was LongCat-2.0, a trillion-parameter model from Meituan, the Chinese food delivery company. They just open-sourced it under MIT license.

The benchmark numbers are real. LongCat-2.0 scores 59.5 on SWE-bench Pro, edging past GPT-5.5's 58.6. It hits 77.3 on SWE-bench Multilingual, matching Claude Opus 4.6 at 77.8. For a model from a company whose main product is delivering hotpot, that's a statement.

But the benchmark headline obscures the actual story. What matters is how they built it and what it costs.

The 50,000-card question

Meituan trained LongCat-2.0 entirely on domestic Chinese compute. No NVIDIA hardware. Their cluster uses AI ASIC superpods, and they scaled to 50,000 cards without a single irrecoverable loss spike. They achieved a 70% reduction in monthly fault rates through elastic scaling and auto-recovery. Steady-state throughput exceeded 1 trillion tokens per day.

This is not a toy experiment. It's a production training run on alternative silicon that produced a model competitive with the best Western labs. The implications for compute independence are hard to overstate. If Chinese companies can train frontier-scale models on domestic hardware, the export control strategy gets a lot more complicated.

What 1.6 trillion parameters actually do

The model activates roughly 48 billion parameters per token, shifting between 33B and 56B depending on task complexity. Simple tokens get zero computation. Complex reasoning triggers the full expert ensemble. This dynamic activation is what keeps inference costs manageable despite the massive total parameter count.

The architecture uses something called MOPD, Multi-Teacher On-Policy Distillation. Three specialized expert groups handle different workloads: Agent experts for tool use and self-correction, Reasoning experts for multi-hop logic, and Interaction experts for instruction following and hallucination suppression. Rather than training a monolithic model and hoping it handles everything, Meituan trained specialists and fused them into a unified system.

LongCat Sparse Attention enables the million-token context window by shifting attention complexity from quadratic to linear. Instead of attending to every token, the model selects key information intelligently. This is not new in concept, but making it work at 1.6T scale with consistent performance across the full context is an engineering feat.

The cost math

At $0.75 per million input tokens and $2.95 per million output, LongCat-2.0 undercuts every proprietary model at similar capability. Cache hits drop to $0.015 per million tokens. For teams running agentic coding workflows that burn through millions of tokens per session, the economics shift dramatically.

The pricing positions it between DeepSeek V4-Flash (cheaper but smaller) and GPT-5.5 (more expensive but also stronger on some tasks). For repository-level code understanding and multi-file refactoring, the 1M context window at this price point is genuinely hard to beat.

What the benchmarks do not tell you

SWE-bench Pro measures something narrow: can the model fix real GitHub issues? It does not measure code quality, maintainability, or whether the fix introduces new bugs. A 59.5 score means it solved 59.5% of benchmark issues, not that it writes 59.5% as well as GPT-5.5.

The weights are not publicly available yet. The model is accessible via API and OpenRouter, but self-hosting requires the full 1.6T parameter set. Even with dynamic activation, that is a lot of hardware. Meituan says weights are "coming soon" on Hugging Face, but until they drop, the open-source claim has an asterisk.

And there is the usual benchmark inflation concern. Labs optimize for specific benchmarks, and SWE-bench has become the de facto standard for coding models. Scores on other tasks, real-world developer experience, and long-horizon agentic performance tell a different story. LongCat-2.0 looks strong on paper. Whether it holds up in production coding workflows is an open question.

The bigger picture

Chinese AI labs are closing the gap faster than most people expected. DeepSeek V4-Pro matched Gemini 3.1 Pro on SWE-Bench Verified at one-thirtieth the cost. MiniMax M3 brought multimodal frontier coding at $0.30 per million input tokens. Now LongCat-2.0 joins the party with trillion-scale open weights under MIT license.

The pattern is clear: Chinese companies are not just matching Western benchmarks. They are doing it cheaper, releasing the weights, and building on hardware that does not depend on Silicon Valley supply chains. Whether you view this as healthy competition or a strategic concern depends on which side of the Pacific you sit on. Either way, the era of a single company dominating AI is over.

Sources