Quick Verdict
LongCat-2.0 is Meituan's open-sourced Mixture-of-Experts coding model — 1.6 trillion total parameters, roughly 48 billion active per token, and a native 1-million-token context window, released June 30, 2026 under a permissive MIT license. The twist is that most of the internet had already been using it for two months without knowing it: it ran anonymously on OpenRouter under the codename "Owl Alpha," quietly climbing to the top of several agent-usage leaderboards before Meituan revealed itself as the team behind it.
On Meituan's own benchmarks, LongCat-2.0 scores 59.5 on SWE-bench Pro versus GPT-5.5's 58.6 — a real but narrow edge on one test, not a clean sweep. The bigger claim is architectural: Meituan says this is the first trillion-parameter-scale model trained and served entirely on domestic Chinese AI chips, with no Nvidia hardware anywhere in the pipeline. This review is based on Meituan's launch materials, the model's own GitHub and Hugging Face pages, and independent reporting from outlets including VentureBeat and TestingCatalog — not yet our own hands-on testing, since the model became publicly attributable only days before this article was written.
Our early take: 7.9/10 — a genuinely capable open-weight agentic coder with unusually strong real-world usage evidence behind it, docked because the benchmark wins are vendor-reported and the downloadable weights aren't actually available yet.
What Is LongCat-2.0?
LongCat-2.0 is a large language model built by Meituan — yes, the Chinese food-delivery and local-services giant, which has quietly been building a serious AI research effort under the LongCat brand (previously known for LongCat-Flash-Chat, LongCat-Video, LongCat-Image, and other projects). This release is a 1.6-trillion-parameter Mixture-of-Experts (MoE) system that dynamically activates only around 33–56 billion parameters per token, purpose-built for agentic coding: tool use, multi-step execution, long-horizon reasoning, and working across an entire codebase rather than answering isolated prompts.
It ships open source under the MIT license, with pretraining across more than 35 trillion tokens. One important caveat: at launch, the GitHub and Hugging Face pages are live and documented, but the actual downloadable model weights are listed as "coming soon" — so the practical way to use LongCat-2.0 right now is through Meituan's hosted API or OpenRouter, not a local download.
The "Owl Alpha" Stealth Launch
The most unusual part of this release isn't the spec sheet — it's how the model already had a track record before anyone knew its name. For roughly two months before the official reveal, LongCat-2.0 ran anonymously on OpenRouter under the codename "Owl Alpha." During that stealth run it became one of the platform's most-used models, reportedly handling around 10.1 trillion tokens a month (about 559 billion tokens a day) with month-over-month volume growth Meituan put at 242%. OpenRouter's own listings had it ranked in the global top 3 by daily volume, and reportedly #1 on Hermes Agent, #2 on Claude Code usage, and #3 on OpenClaw usage — all while developers had no idea which lab was behind it.
On June 30, 2026, Meituan's official LongCat account confirmed the connection directly: "Owl Alpha on OpenRouter — that's us." That's a meaningfully different launch story than most model releases, where benchmark charts and marketing come first and real usage follows. Here, heavy developer usage came first, and the reveal came after.
How LongCat-2.0 Is Built
Running a 1.6-trillion-parameter model at OpenRouter scale without bankrupting Meituan comes down to a few specific architectural choices, rather than just "more parameters."
- LongCat Sparse Attention (LSA): Standard attention compares every token against every other token, so cost grows quadratically as context length increases. LSA instead selects only the most relevant tokens to attend to, which is what makes the full 1-million-token context window practical instead of prohibitively slow — letting the model pull one relevant function out of a huge codebase without re-processing everything else.
- Zero-Compute Experts: In most MoE models, every token still pays a baseline compute cost. LongCat-2.0 routes simple, low-difficulty tokens through near-empty subnetworks that consume almost no compute, while harder tokens get routed to more, heavier experts — which is why active parameters flex between roughly 33B and 56B per token instead of staying fixed.
- Three specialized expert groups (MOPD): Training is split across Agent Experts (tool use, self-correction), Reasoning Experts (multi-hop reasoning, adaptive computation), and Interaction Experts (instruction-following, hallucination suppression), fused together through a method Meituan calls Multi-Teacher On-Policy Distillation. A gating network then routes each task to whichever expert group fits it best.
LongCat-2.0 Pricing: API Costs
LongCat-2.0 is priced per token through Meituan's LongCat API Platform, with a limited-time launch promotion that undercuts the standard rate significantly.
| Access method | Price | Notes |
|---|---|---|
| LongCat API — standard rate | $0.75 / 1M input tokens, $2.95 / 1M output tokens | Listed post-promo pricing |
| LongCat API — launch promo | $0.30 / 1M input tokens, $1.20 / 1M output tokens | Limited-time introductory rate as of launch |
| Cached context reads | Free (standard) / $0.015 per 1M (some listings) | Cache hits don't count against token-pack usage |
| OpenRouter | Same underlying model, aggregator routing | Previously listed anonymously as "Owl Alpha" |
For comparison, GPT-5.5's API pricing runs roughly $5 per million input tokens and $30 per million output tokens — meaning LongCat-2.0's promotional rate is on the order of 15–25x cheaper per token on paper, which is the core of Meituan's pitch for high-volume agentic workloads.
Try LongCat-2.0 through the aggregator that hosted it as "Owl Alpha"
See LongCat-2.0 on OpenRouter →LongCat-2.0 vs GPT-5.5 and GLM-5.2: Benchmarks
All figures below are Meituan's own vendor-reported numbers, cross-referenced against independent coverage from VentureBeat and TestingCatalog. As with any self-reported benchmark, treat these as a starting point, not a substitute for testing your own workload — and note that the "beats GPT-5.5" framing rests on a single test, decided by under a point.
| Benchmark | LongCat-2.0 | GPT-5.5 | GLM-5.2 |
|---|---|---|---|
| SWE-bench Pro (real-world coding) | 59.5 | 58.6 | 62.1 |
| Terminal-Bench 2.1 (agentic CLI tasks) | 70.8 | Not independently confirmed | ~81.0 |
| SWE-bench Multilingual | 77.3 | — | — |
| BrowseComp (deep web research) | 79.9 | — | — |
| Context window | 1,000,000 tokens | Varies by tier | 1,048,576 tokens |
| License | MIT (weights pending) | Closed/proprietary | MIT (weights available) |
The honest read: LongCat-2.0 lands in genuinely near-frontier territory for agentic coding and web research, which is impressive for an open-weight release, but it isn't a clean win over every open competitor either — GLM-5.2, which we reviewed in June, still reports a higher Terminal-Bench score and a slightly higher SWE-bench Pro number. LongCat-2.0's strongest, least-ambiguous edge shows up in agentic search and browsing tasks (BrowseComp, RWSearch), not raw coding benchmarks alone.
Key Features
- 1M-token native context window: Enough to hold an entire mid-size repository or a long multi-turn agent session in memory at once, made practical by LongCat Sparse Attention rather than brute-force scaling.
- Trained without Nvidia hardware: Meituan says the full training and inference pipeline ran on a cluster of more than 50,000 domestic Chinese AI chips — the first trillion-parameter-scale model, per Meituan, built entirely on non-Nvidia, non-Google-TPU infrastructure.
- Dynamic compute routing: Zero-Compute Experts mean simple tokens cost almost nothing to process while hard tokens get proportionally more compute, rather than a flat cost per token regardless of difficulty.
- Proven agentic usage before the reveal: Two months of heavy, anonymous real-world OpenRouter usage as "Owl Alpha" is a stronger signal of practical usefulness than a typical launch-day benchmark chart.
- OpenAI- and Anthropic-compatible API formats: Supports drop-in style integration with tools like Claude Code, OpenClaw, OpenCode, Kilo Code, and Codex-style agent workflows.
How to Access LongCat-2.0
LongCat-2.0 is available today through hosted access, though the fully self-hostable open-weight download isn't live yet:
- LongCat API Platform: Pay-per-token access directly from Meituan at the promotional rate above; check longcat.chat for current pricing once the launch promo ends.
- OpenRouter: The same model previously listed anonymously as "Owl Alpha," now attributed to Meituan directly, with OpenAI- and Anthropic-compatible API formats.
- Hugging Face and GitHub: Both pages are public under the MIT license and documented, but as of launch the downloadable weights are listed as "coming soon" rather than immediately available.
- Agent tooling integrations: Reported compatibility with Claude Code, OpenClaw, OpenCode, and Kilo Code workflows via standard API formats.
Check the official model card and pending weight release
View LongCat-2.0 on Hugging Face →Pros and Cons
| Pros | Cons |
|---|---|
| Two months of real, heavy anonymous usage on OpenRouter before launch — unusually strong pre-launch validation | Downloadable open weights are not yet available; GitHub/Hugging Face list them as "coming soon" |
| Aggressive promotional pricing, far cheaper per token than GPT-5.5 | The "beats GPT-5.5" claim rests on a single benchmark, decided by under a point |
| 1M-token context window, purpose-built for agentic coding and long-horizon tasks | All benchmark figures are vendor-reported by Meituan, not independently verified |
| Notable architecture (LSA, Zero-Compute Experts) that keeps a 1.6T model genuinely affordable to run | Trails GLM-5.2, another open-weight Chinese model, on Terminal-Bench and SWE-bench Pro |
| First trillion-parameter-scale model reportedly trained entirely without Nvidia hardware | Hosted API access routes data through Meituan's infrastructure unless you wait for self-hostable weights |
Who Should Use LongCat-2.0?
Worth testing now if you are:
- Running high-volume agentic coding or research workloads where the 1M-token context and aggressive promotional pricing meaningfully cut API spend versus GPT-5.5.
- Building agent tooling that already talks to Claude Code, OpenClaw, or Codex-style APIs — LongCat-2.0's compatible API formats make it a low-friction model to test alongside what you're already using.
- Curious about the no-Nvidia training story for its own sake — the compute-independence angle matters beyond just this one model's benchmark scores.
Probably wait if you:
- Need self-hostable open weights today — LongCat-2.0's weights aren't downloadable yet, unlike GLM-5.2's.
- Want independently verified benchmarks rather than vendor-reported figures before committing production workloads.
- Have strict requirements against routing data through China-based infrastructure and can't wait for a self-hosted option.
Frequently Asked Questions
Is LongCat-2.0 free?
The model isn't free to run at scale — you pay per token through Meituan's LongCat API Platform or OpenRouter, currently at a promotional rate of $0.30/$1.20 per million input/output tokens. The MIT-licensed weights will be free to self-host once Meituan releases them, but as of launch they're listed as "coming soon" rather than downloadable.
What was "Owl Alpha"?
Owl Alpha was the anonymous codename LongCat-2.0 ran under on OpenRouter for roughly two months before Meituan's official reveal on June 30, 2026. During that stealth period it became one of the platform's most-used models by volume, reportedly handling around 10.1 trillion tokens a month, before Meituan confirmed it was the team behind it.
Is LongCat-2.0 actually better than GPT-5.5?
On SWE-bench Pro, the benchmark Meituan highlighted at launch, LongCat-2.0 scored slightly higher (59.5 vs 58.6) — a narrow, vendor-reported margin rather than a decisive win. It also trails another open-weight model, GLM-5.2, on Terminal-Bench and SWE-bench Pro. Whether it's "better" for your workload depends on the specific task, and we haven't yet run our own side-by-side testing.
Can I self-host LongCat-2.0?
Not yet. The GitHub and Hugging Face pages are public under the MIT license, but at launch the downloadable weights are listed as "coming soon." Right now, the only ways to use LongCat-2.0 are through Meituan's hosted API or OpenRouter.
Was LongCat-2.0 really trained without Nvidia chips?
That's Meituan's claim: the model was trained and is served on a cluster of more than 50,000 domestic Chinese AI chips, with no Nvidia GPUs involved — which the company says makes it the first trillion-parameter-scale model built entirely on non-Nvidia infrastructure. This hasn't been independently audited by a third party as of this writing.
Final Verdict
LongCat-2.0 is a legitimately interesting release — not just for the benchmark numbers, which are real but modest, but for the way it arrived: two months of unglamorous, heavy real-world usage on OpenRouter before anyone knew whose model it was. That's a rarer and arguably more convincing form of proof than a typical launch-day chart, and paired with genuinely novel architecture (LongCat Sparse Attention, Zero-Compute Experts) and a no-Nvidia training story, it's a release worth paying attention to.
The honest caveats: every benchmark here is self-reported by Meituan, the "beats GPT-5.5" framing is a one-test, sub-one-point margin, and the open weights you'd need to self-host it aren't actually available yet. This article reflects launch-week specs and reporting, not our own extended testing — we'll revisit with hands-on notes once the weights ship and we've run it against real workloads.
Early rating: 7.9/10 — a promising, aggressively priced open-weight coding model with unusually credible real-world usage behind it, held back only by unverified benchmarks and weights that aren't downloadable yet.