Remember Owl Alpha? It was a mysterious stealth model that quietly climbed the OpenRouter and Kilo leaderboards for weeks while everyone argued about which lab it came from.
Some said it was the latest GLM model from Zai. Some were sure it was a Kimi variant from MoonshotAI. Nobody — nobody — guessed it was a 1.6-trillion-parameter monster from a food delivery company.
At the tail end of Q2, it was revealed that Owl Alpha was LongCat-2.0 from Meituan. There was a ton of buzz around the model, and a good amount of the 10T+ tokens consumed during the stealth period were from Kilo users.
If you were one of the many Kilo Coders who fell for Owl Alpha before the reveal, we’ve got good news: LongCat 2.0 is free on Kilo for a limited time. No limits, just put it to use.
LongCat 2.0
Short version: LongCat 2.0 is a 1.6-trillion-parameter Mixture-of-Experts model that activates about 48 billion parameters per token, built specifically for agentic coding. It ships on Hugging Face under a plain MIT license, which is great news for anybody moving toward local AI in production.
Meituan has said that the new LongCat is the industry’s first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic cluster. Well, whatever they did, it worked. The architecture is doing some clever things to keep that scale affordable. LSA sparse attention extends the long context window without the usual quadratic cost blowup, Zero-Compute Experts keep per-token activation between roughly 33B and 56B, and a routing scheme splits work across separate Agent, Reasoning, and Interaction expert groups — so you’re not burning cycles on experts that have nothing to do with the task in front of them.
But does it actually code? How much does it cost?
The Owl Alpha era answered that better than any benchmark table could. Running fully anonymous, with no brand halo or geographical info attached, it won over a lot of devs. On paper it holds up too. Meituan reports it beats a Claude Opus reference on IFEval (90.0 vs 86.0) and IMO-AnswerBench (81.8 vs 75.3), while trailing on SWE-bench Pro (59.5 vs 69.2).
Per-million-token pricing is a trap. What you actually pay depends on how many tokens a model burns to finish the job — verbosity, reasoning overhead, retries, all of it. So look at cost per completed task instead.
On Artificial Analysis’s weighted cost-per-task metric, LongCat 2.0 comes in at $0.12 per Intelligence Index task. DeepSeek’s latest flagship, V4 Pro 0813, lands at $0.25 per task — more than double. Same open-weight class, same 1M-context ambitions, roughly the same architecture on paper (both 1.6T MoE), and LongCat does the job for less than half the spend per task. Point being: this isn’t “cheap because it’s worse.” It’s affordable because it’s efficient about how it gets to an answer, and that efficiency compounds hard once an agent is chewing through a real workload turn after turn.
And again, for a limited time in Kilo, that cost-per-task number is zero.
If you’ve read about the new Meta(flow), you already know how we think about building now. Kilo handles the plan-and-build side, agent modes map cleanly onto how a real pipeline takes shape, and the Anaconda Platform takes it to production. Plan mode sketches the graph. Code mode implements each step. Debug mode catches what’s broken. Review mode closes the loop before anything ships.
And remember: you don’t have to run the same model in every mode or agent. That’s the whole point of Kilo being model-agnostic. And LongCat 2.0 is an unusually good fit for the modes where it counts.
There’s never been a better time for model freedom
Let’s be honest about the current standings. Look at our leaderboard and the top of the chart is a familiar crowd: OpenAI’s GPT-5.6 Sol and GPT-5.5 are running one-two on KiloBench, with SpaceXAI’s Grok 4.6 right behind them at number three — and those same models are near the top of our usage numbers too. When something is that good, developers reach for it, and we’re not going to pretend otherwise. Sol, Luna, and Grok earn their spots.
But “what’s winning this week” and “what’s worth your attention” aren’t the same list. Kilo has always stood for model choice and model freedom, consistently supporting more than 500 of the top models for builders. Part of that commitment is making sure a genuinely strong performer doesn’t get lost just because it came out of a food delivery company instead of a lab everyone already watches.
LongCat 2.0 is exactly that kind of model: it beat name-brand competition while nobody knew who made it, and it does the job for less per task than DeepSeek’s flagship. You deserve to know it’s there — and to try it without a bill attached — before you default to the usual top of the chart.
Let’s go!
LongCat 2.0 is available on Kilo right now, free for a limited time.
Like everything on Kilo, it works with our model-agnostic setup. No upcharge, no lock-in, no games.
If you tried Owl Alpha and quietly wondered where it went, it didn’t go anywhere. It just got a name, a license, and a spot in your agent lineup. Wire it into your flow, hand it a gnarly step, and see why it topped the charts before anyone knew who made it.
Fire it up in Kilo today — the meter’s off while the promo lasts.
This model is not totally free :)
It it terrible. To test it, I asked it to fix some broken code. Not only has it failed to fix anything, it has made it even worse. Now it is generating errors, and the POS cannot even get out of that.