Meituan Just Open-Sourced a 1.6 Trillion Parameter AI — Built on Chinese Chips
The company most people know for delivering hotpot at midnight just released one of the most ambitious open-source AI models of 2026. On June 30, Meituan shipped LongCat-2.0 — a 1.6 trillion parameter model that runs on Chinese hardware, holds a million tokens in its head at once, and is already among the top-3 most-used models on OpenRouter.
If you've been watching the open-source AI race, this one is worth paying attention to.
What LongCat-2.0 Actually Is
LongCat-2.0 is a Mixture of Experts (MoE) model. It has 1.6 trillion total parameters, but only around 48 billion activate for any given input. That's how it stays fast enough to be practical despite its scale.
The headline number that matters most: a native 1 million token context window. That's not a hack or an approximation — it's baked into the architecture from the start. For comparison, most models that advertise long contexts either approximate it or show significant quality degradation past a few hundred thousand tokens.
On benchmarks, LongCat-2.0 scores 59.5 on SWE-bench Pro and 77.3 on SWE-bench Multilingual — the coding agent benchmarks the industry uses to measure how well models can actually fix bugs in real repositories. Those are competitive numbers by any standard.
Why "Built on Chinese Chips" Is the Real Story
LongCat-2.0 was trained on a cluster of 50,000 domestic Chinese GPUs — no NVIDIA H100s, no A100s. That's significant beyond the geopolitics.
The AI industry has quietly assumed that cutting-edge model training requires US-origin hardware. Export restrictions have been treated as a hard ceiling on what Chinese labs can build at scale. LongCat-2.0 is evidence that the ceiling is higher than assumed, and that domestic Chinese compute has matured enough to run the full training pipeline — not just inference — on models of this size.
A New Architecture for Long-Context Agent Work
The model introduces several architectural choices worth noting:
- LSA (Longitudinal Sparse Attention): Designed to handle very long sequences without the usual quadratic memory cost.
- Zero-compute experts: Experts in the MoE that contribute to routing decisions while consuming near-zero FLOPs during inference.
- MOPD multi-expert fusion: Three specialized expert groups — Agent, Reasoning, and Interaction — handle different task types in parallel.
This isn't just throwing more parameters at the problem. It's a specific architectural bet that long-context, multi-task agent work requires different design choices than pure chat or single-turn coding.
Already Used by Millions
LongCat-2.0 isn't just a research release. It's live on OpenRouter and at longcat.ai, and it climbed into the top 3 globally on OpenRouter by usage before the formal open-source release even dropped.
That's a remarkable adoption number for a model that most of the Western AI world hasn't heard of. The preview launched first, so there's real-world usage data behind these numbers — not just benchmark figures in a paper.
What This Means If You Use OpenClaw
The rise of capable open-source models is directly relevant to anyone building or running AI agents.
When OpenClaw launched, the practical choices for capable agents were a small set of frontier models — most of them closed-source and US-based. LongCat-2.0 is part of a wave of models changing that equation. More capable open-source options mean more flexibility: where you run your agents, how you price them, and what you can customize.
The 1M context window is especially relevant for agents specifically. Agents aren't answering one question — they're tracking tool calls, memory, prior task outputs, and ongoing session context simultaneously. Models that can genuinely hold more in context aren't just faster; they're qualitatively better at multi-step work.
As open-source models continue closing the gap with proprietary ones, the agent ecosystem gets more interesting — more models to route between, more options for cost-sensitive workloads, and more community development on shared foundations.
That's the world OpenClaw is built for.