There's a reliable way to tell whether an AI model is actually good versus just well-marketed: watch what developers reach for when nobody is selling it to them.
That's what happened on OpenRouter starting in early May.
An anonymously listed model appeared on the platform under the name "Owl Alpha." No announcement. No company behind it. No LinkedIn post from a founder announcing "we're thrilled to." Just a model in the registry that developers could try.
They tried it. They kept using it. They told other developers.
By late June, Owl Alpha had accumulated 10.1 trillion monthly tokens processed — ranking first globally on the Hermes Agent coding workspace by call volume, with +242% month-over-month growth. That's not benchmark theater. That's developers repeatedly reaching for something because it works.
Then on June 30, the curtain dropped. Meituan — yes, the Chinese food-delivery-turned-tech-conglomerate — revealed that Owl Alpha was their model: LongCat-2.0-Preview. And they simultaneously open-sourced the full weights under an MIT license.
What the Numbers Actually Look Like
LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts model, but only about 48 billion parameters are active at any point during inference — which is what makes it practical to run and fast enough to be useful. It carries a native 1-million-token context window.
On SWE-bench Pro — the benchmark that tests models on real GitHub issues pulled from production codebases, not synthetic exercises — LongCat-2.0 scores 59.5. GPT-5.5 scores 58.6. Gemini 3.1 Pro comes in at 54.2. It's not at the very top of the frontier, but it's ahead of models that cost dramatically more to run.
The pricing is where the story gets interesting for anyone running budget-constrained workflows:
| Model | Input (per 1M tokens) | Output (per 1M tokens) | |---|---|---| | LongCat-2.0 (standard) | $0.75 | $2.95 | | LongCat-2.0 (launch promo) | $0.30 | $1.20 | | Cached context reads | Free | — | | GPT-5.5 | $5.00 | $30.00 |
At output pricing, that's a 10x gap in standard rates and a 25x gap during the promo. If you're running a coding agent, a document Q&A pipeline, or any workflow that processes significant token volume, this is a budget line item that changes what you can afford to build. Meituan also offers token packs — roughly $60 for 1 billion tokens — which makes high-volume use cases even cheaper.
The Part About Chinese Chips
LongCat-2.0 is the first trillion-parameter model trained end-to-end on Chinese-made chips. No NVIDIA hardware involved, start to finish. Meituan used a 50,000-chip cluster of domestic ASICs — technology from Huawei's Atlas line — for both pre-training and inference.
The practical implication for your org isn't geopolitical. It's economic.
The US export control strategy was premised on the idea that restricting access to high-end NVIDIA chips would meaningfully slow China's AI development. The reveal of LongCat-2.0 — a near-frontier model trained without those chips, that developers organically chose over everything else during a two-month anonymous trial — is a pretty direct answer to that premise.
What this means downstream: the cost competition in AI APIs is not going to ease. Chinese labs have now demonstrated they can train frontier-class models on domestic hardware. The models will keep coming, they'll keep being priced aggressively, and that's genuinely good for any org that buys AI services rather than sells them.
You Might Already Be Using It
If your team is routing AI calls through OpenRouter, there's a non-trivial chance some of your requests were hitting Owl Alpha before it was unmasked. OpenRouter's anonymous model listings are deliberately obscured during preview phases, and Owl Alpha was one of the fastest-growing models on the platform for months.
That's not a concern — it's actually useful information. If any of your existing workflows were performing well and cost-efficient during that period, you've already stress-tested LongCat-2.0 in production. The model that developers were organically choosing at scale was this one.
The Self-Hosting Path
The weights are on Hugging Face under meituan-longcat. Under MIT license, you can download, modify, fine-tune, and run the model in your own infrastructure without any data going anywhere external.
Running a 1.6T parameter model in full precision requires serious hardware — we're talking multi-GPU clusters or well-provisioned cloud instances. But the MoE architecture means you're only activating ~48B parameters per forward pass, which brings the actual compute requirements closer to a mid-range dense model. Teams with data residency requirements or compliance reasons to keep data on-premise have a real self-hosting path here.
For most small teams, the API is the right starting point. Get a result first, then decide if the self-hosting case is worth building.
Where This Fits Your Workflow
Three use cases where LongCat-2.0's cost and context window make the most difference:
High-volume document processing. If you're running grant databases, policy archives, contracts, or large report backlogs through an AI pipeline, the 1-million-token context and free cached reads change the economics substantially. Feed the whole document set in one call rather than chunking it into dozens of expensive requests.
Coding agents on a budget. The SWE-bench numbers are real-world, not synthetic. If you're building internal tools, automating data pipelines, or running maintenance on legacy code, the performance is near-frontier at a fraction of the cost.
Agentic workflows with long chains. LongCat-2.0 was specifically built for multi-step agentic tasks. If you're stringing together five or ten API calls in a chain, cutting output costs by 10x compounds across every step.
What to Do This Week
If you're already using OpenAI or another premium provider for any of these use cases, the test is simple: run your existing prompts through LongCat-2.0 via OpenRouter or the Meituan API, compare output quality and cost side by side, and make a call with actual data.
OpenRouter setup takes about 20 minutes if you already have API calls in your codebase — LongCat-2.0 uses the OpenAI-compatible API format, so you're changing a base URL and a model name, not rewriting anything.
The stealth period tells you something. Developers don't keep using a model 10 trillion times a month because of hype. They use it because it works. Now that it's named, licensed, and priced, there's no reason not to evaluate it for your own stack.
This is the kind of model evaluation we help teams run in a half-day — scope the use case, test against your actual workload, make a concrete recommendation. No six-month roadmap required.