LongCat-2.0: China Just Dropped a 1.6 Trillion Parameter Model You Can Use for Free

LongCat-2.0 - China Just Dropped a 1.6 Trillion Parameter Model You Can Use for Free

Links & Resources

Meituan just open-sourced a 1.6 trillion parameter AI model. It's called LongCat-2.0, it's built for agentic coding, and it's been quietly dominating global developer charts for two months before anyone knew what it was.

The scale sounds absurd. But the architecture is smarter than the headline. Only a tiny slice of those 1.6 trillion parameters fires for each token - 48 billion on average - so it runs fast and cheap despite being nearly triple the size of DeepSeek-V3. And the part that changes the conversation: the entire model was trained on ~50,000 Chinese-made chips with zero Nvidia involvement. The biggest training run ever done without them.

Released June 30, 2026. MIT license. Fully open source.

What Is LongCat-2.0

LongCat-2.0 is a Mixture-of-Experts (MoE) large language model from Meituan, China's delivery and local-services giant that has been quietly building one of the country's most serious AI labs. It targets agentic coding - not one-shot chat answers, but multi-step code understanding, generation, tool calling, and autonomous execution inside agent workflows.

The Owl Alpha Backstory

Before it had a name, LongCat-2.0 ran anonymously on OpenRouter for two months under the codename "Owl Alpha." During that stealth run, it became one of the most-used models on the platform:

Developers were reaching for this model heavily before anyone knew Meituan was behind it. It climbed to the top of multiple agent leaderboards and sat near the top on Claude Code deployments - all while being completely anonymous. That's a stronger signal than any launch-day benchmark chart.

On June 30, Meituan pulled the mask off: Owl Alpha was LongCat-2.0 the entire time.

The Architecture - Why 1.6T Doesn't Mean Slow

A 1.6 trillion parameter model sounds like it should cost a fortune to run. Three architectural innovations make it cheap instead.

1. LongCat Sparse Attention (LSA)

Standard attention compares every token to every other token - cost grows quadratically with context length. LSA selects only the most relevant tokens to attend to, dropping the scaling closer to linear. That's what makes the full 1-million-token context window practical rather than ruinously expensive. The model can pull a specific function out of an entire codebase without re-reading everything.

2. Zero-Compute Experts

In most MoE models, every token still pays a baseline compute cost. Zero-Compute Experts let simple tokens route through near-empty subnetworks that consume almost no compute, while hard tokens get more experts assigned. This is why the active parameter count flexes between 33B and 56B per token instead of staying fixed - easy boilerplate is cheap, tricky logic gets the full horsepower.

3. MOPD - Three Specialized Expert Groups

LongCat-2.0 splits its training into three expert families fused through Multi-Teacher On-Policy Distill (MOPD):

A gating network routes each task to the expert family best suited for it at inference time, rather than blending everything into one averaged model.

Benchmarks

Every number below is vendor-reported by Meituan and not independently verified. The "beats GPT-5.5" headline is true on exactly one benchmark by under a single point. Treat these as a claimed ceiling.

The honest read: LongCat-2.0 lands in near-frontier territory for agentic coding, which is genuinely impressive for an open-weight model. The GPT-5.5 comparison is a marketing-friendly framing of a statistical tie on one test. The browsing and search scores are the standout numbers and line up with the heavy agent usage it saw as Owl Alpha.

On broad general-agent benchmarks like FORTE and BrowseComp, it generally trails premium closed-source systems like Claude Opus 4.8 - but that's expected for an open-weight model priced at a fraction of the cost.

Trained Without Nvidia - The Bigger Story

This is the detail drawing the most attention, and it should be.

Meituan says LongCat-2.0 was trained and served on a cluster of 50,000+ Chinese-made ASICs - likely Huawei Ascend 910C chips based on community analysis - with zero Nvidia GPUs in the loop. Meituan claims this makes it the first trillion-parameter model to complete full-process training and inference entirely on domestic compute.

For context: even DeepSeek-V4-Pro used Chinese chips only for inference, relying on Nvidia for pre-training. LongCat-2.0 used domestic hardware for both.

This matters because it proves frontier-scale training is now technically viable on Chinese silicon. While the US tightens chip export controls, China's labs keep shipping:

Each release pushes AI forward while keeping access open for everyone. The pattern is unmistakable.

Pricing - An Order of Magnitude Cheaper

This is where LongCat-2.0 gets aggressive.

The promotional pricing is roughly 17x cheaper on input and 25x cheaper on output than GPT-5.5. For heavy agent workloads burning millions of tokens a day, that's a massive gap.

Free context-cache hits are the cherry on top. For agentic coding where the same codebase context gets reused across many calls, cached tokens cost literally nothing.

How to Use LongCat-2.0 Right Now

Option 1 - OpenRouter (Easiest)

The model is live on OpenRouter - the same platform where it ran as Owl Alpha. You can access it through any tool that supports OpenRouter's API (Claude Code, Cursor, etc.).

Option 2 - LongCat Platform

Meituan's own hosted platform at longcat.chat provides direct API access with the promotional pricing.

Option 3 - API Integration

# OpenRouter API example
import openai

client = openai.OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="your-openrouter-key"
)

response = client.chat.completions.create(
    model="meituan/longcat-2.0",
    messages=[
        {"role": "user", "content": "Refactor this function to use async/await..."}
    ]
)

Option 4 - Self-Hosted (Coming Soon)

The Hugging Face model card and GitHub repo are live, but the downloadable weights are listed as "coming soon" at launch. When released, the MIT license allows full commercial self-hosting.

API Documentation

Full docs at longcatai.org/docs

What LongCat-2.0 Is Built For

This isn't a general chatbot. It's built specifically for:

Agentic Coding - Multi-step code generation, debugging, and refactoring inside agent loops. The 1M context window means it can hold an entire codebase in memory while working on specific files.

Tool Calling - Native tool-use capability optimized by the Agent Expert group. The model was designed from the ground up to call APIs, execute commands, and chain tool outputs.

Long Multi-Step Tasks - The combination of sparse attention + 1M context + agent experts means it can sustain coherent execution across hundreds of steps without losing track of earlier context.

Repository-Scale Work - Feed it an entire repo. The sparse attention architecture means it doesn't choke on large codebases the way dense models do.

What to Watch For

Benchmarks are self-reported. Every number comes from Meituan's internal testing. Independent verification on third-party leaderboards is pending. The "beats GPT-5.5" claim is a sub-1-point margin on a single benchmark.

Weights aren't downloadable yet. The MIT license and open-source promise are real, but you can't actually download and self-host the model at launch. API access only for now.

Broad general knowledge. LongCat-2.0 is optimized for agentic coding. If your use case is creative writing, general Q&A, or non-coding tasks, frontier closed-source models will likely still outperform it.