LongCat-2.0

Meituan is the company that delivers your dumplings in Beijing. On June 29, 2026, its LongCat team open-sourced LongCat-2.0, a 1.6-trillion-parameter agentic model under the MIT license, with a native 1M-token context window. Two more headlines ride along: the model had already spent two months anonymously topping OpenRouter's agent leaderboards as "Owl Alpha," and Meituan says the whole thing was trained and served on domestic Chinese AI chips, no NVIDIA hardware involved. Here is what shipped, the numbers with their asterisks, and what you can actually do with it today.

What Meituan shipped

LongCat-2.0 is a sparse Mixture-of-Experts (MoE) model: 1.6 trillion total parameters, with roughly 48 billion active per token. An MoE splits its feed-forward layers into many small "expert" subnetworks and routes each token through only a few of them, so you pay storage for all 1.6T but compute for about 48B, the same bargain behind DeepSeek-V4 and GLM-5.2. LongCat goes one step further: zero-computation experts and a PID controller let the activation float between about 33B and 56B per token depending on how hard the token is, per the official README.

The release is complete in the way we wish every open drop were: BF16 weights, FP8 and INT8 checkpoints, SGLang deployment guides for both GPUs and NPUs, OpenAI- and Anthropic-compatible API endpoints, a thinking-mode toggle, and MIT for all of it. Pretraining ran on more than 35 trillion tokens, including hundreds of billions of tokens of 1M-context data, and Meituan claims the run finished with no rollbacks or loss spikes. Maximum output is 131,072 tokens.

This is not Meituan's first model. LongCat-Flash, a 560B MoE from 2025 (tech report), introduced the zero-computation-expert trick, and the Flash-Thinking, Flash-Omni, and Flash-Lite variants followed. LongCat-2.0 is the trillion-scale successor, and it arrives in the same June wave that gave us GLM-5.2 and Kimi K2.7 Code.

Two months undercover as Owl Alpha

The best part of the launch story is that the market reviewed this model before anyone knew what it was. LongCat-2.0 ran anonymously on OpenRouter as "Owl Alpha" for about two months. During that stealth run it ranked first by monthly call volume on the Hermes Agent workspace, second on Claude Code, and third on OpenClaw, according to Decrypt's report. Community trackers put its throughput above ten trillion tokens a month before the reveal, a figure Meituan itself has not confirmed, so treat it as community-reported.

Why that matters: developers routed real agent workloads to an unbranded model because it performed, not because of a launch blog. That is about as clean a blind test as this industry gets.

The benchmarks, and the asterisk

Meituan's official table compares LongCat-2.0 only against closed frontier models. Values marked with an asterisk are cited from the competitor's own reports; the LongCat numbers are Meituan's and have not been independently reproduced. Notably, there is no official comparison against GLM-5.2 or DeepSeek-V4, so resist the urge to line those up from different tables.

All LongCat figures are Meituan-reported. Asterisked competitor scores are cited from those vendors' own publications; "n/p" means no public score. Source: the official benchmark table.

The honest read: near-frontier, not frontier-beating. On SWE-bench Pro, the agentic coding benchmark people actually watch, 59.5 edges GPT-5.5 (58.6) and clears Gemini 3.1 Pro (54.2), but the newer Claude Opus 4.7 (64.3) and 4.8 (69.2) stay ahead. For an MIT-licensed model you can download, modify, and redistribute, beating any current closed frontier model on any agentic benchmark is the story. Expecting it to beat all of them is not the right bar.

The Chinese-chips claim

Meituan calls LongCat-2.0 the industry's first trillion-parameter model to complete full-process training and inference on domestic "AI ASIC superpods," with reporting citing a cluster of about 50,000 accelerator cards and zero NVIDIA silicon (SiliconANGLE, SCMP).

The part nobody can verify: which chips. Meituan never names the vendor. The community's guess is Huawei's Ascend line, inferred from interconnect and memory clues in the documentation, but that is speculation, not confirmation. What is checkable is the deployment code: the repo ships an NPU serving path alongside the GPU one, which at minimum proves non-NVIDIA inference is a first-class citizen, not a press-release afterthought. If the training claim holds up, the export-control debate we covered in the open-weights export fight just got a large new data point.

Architecture for the curious

Two pieces are genuinely new. First, LongCat Sparse Attention (LSA), Meituan's answer to the sparse-attention design DeepSeek introduced: an indexer picks which past tokens each query actually attends to, and LSA layers three refinements on top (streaming-aware indexing, one index shared across every two layers, and a coarse-to-fine hierarchical pass). The payoff is long-context cost that scales closer to linear, which is what makes a native 1M window usable rather than theoretical.

Second, the on-disk weights are bigger than the headline: alongside the 1.6T MoE parameters sits a 135B N-gram Embedding block, inherited from LongCat-Flash-Lite, which is why the checkpoint folder describes itself as roughly 1.8T parameters. A 3-step multi-token prediction module handles speculative decoding, and post-training distills three separate teacher groups (agent, reasoning, interaction) into one student, a recipe Meituan calls MOPD.

Limitations and gotchas

• You cannot run this at home. Even the FP8 checkpoint is a multi-node deployment; community estimates point at a 16x H20-class server as the floor, and that figure is itself unverified. This is a rent-a-cluster or use-the-API model.
• No Ollama build and no community GGUF exist yet, so there is no one-line local install to paste here.
• Benchmarks are self-reported, and the OpenRouter throughput legend is community-reported. The blind Owl Alpha traffic is the strongest independent signal we have.
• The chip vendor is undisclosed. "Trained on domestic silicon" is a Meituan claim with no third-party audit.
• API pricing (around $0.75 per million input tokens and $2.95 per million output at standard rates, with a launch promo roughly 60 percent below that) comes from aggregator reporting, so confirm current rates on the platform before you build a budget around them.

Who should use it

Teams running agent harnesses (Claude Code-style CLIs, browser agents, long-horizon coding loops) who want an MIT-licensed brain they can eventually self-host, and researchers who want to study trillion-scale sparse attention with full weight access. If you need a model for one GPU, this is not it; look at Qwen 3.6 27B instead. If you need the absolute best agentic scores regardless of license, Claude Opus 4.8 still holds that crown.

Try it in about 10 minutes

The fastest path is the free chat at longcat.ai. For API access, the platform exposes both OpenAI- and Anthropic-compatible endpoints, so existing harnesses connect with a base-URL swap; the README documents the exact setup for Claude Code and friends. If you have serious hardware, the weights are one command away:

# Pull the FP8 checkpoint (bring a multi-node GPU server and patience)
hf download meituan-longcat/LongCat-2.0-FP8

# Deployment recipes for SGLang (GPU and NPU) live in the repo
git clone https://github.com/meituan-longcat/LongCat-2.0

Then point whatever agent harness you already run at the endpoint and give it a long multi-step task. A model tuned for agents deserves a real agent workload, not a haiku request.

Sources and further reading

• LongCat-2.0 model card (Hugging Face)
• Official GitHub repo, benchmark table and deployment guides
• VentureBeat coverage
• Decrypt: the Owl Alpha stealth run
• SCMP: China's biggest model trained on local chips
• MarkTechPost: architecture breakdown
• LongCat-Flash technical report (arXiv)

Tested on: not independently tested. LongCat-2.0 is a roughly 1.8T-parameter checkpoint that needs a multi-node GPU or NPU cluster, beyond our bench. All benchmark figures are Meituan-reported; OpenRouter traffic figures are community-reported and flagged as such. Sources linked above.
Date checked: 2026-07-09