Meituan quietly open-sourced its trillion-parameter model LongCat-2.0: topped the global top 3 in 22 hours. The hottest model on OpenRouter worldwide, the one that processed 10 trillion tokens—turns out it was built by Meituan, a 1.6 trillion-parameter MoE. Trained on 35 trillion tokens, all completed on 50,000 domestic AI chips, without a single NVIDIA GPU. A food delivery company pivoting straight into AI.

💡 What You Will Learn

Meituan quietly open-sourced its trillion-parameter model LongCat-2.0: topped the global top 3 in 22 hours. The hottest model on OpenRouter worldwide, the one that processed 10 trillion tokens—turns o

📜 Table of Contents

• A Food Delivery Company Pivoting to AI
• What's Actually Inside a Trillion-Parameter Model
• The Price Is Genuinely Cheap
• What This Means — Three Takeaways
• Personal Take
• What Do You Think

Meituan Quietly Open-Sources Trillion-Parameter Model LongCat-2.0: Hits Global Top 3 in 22 Hours

The hottest model on OpenRouter globally, the one that processed 10 trillion tokens — turns out it's Meituan's 1.6-trillion-parameter MoE. Trained on 35 trillion tokens, entirely on 50,000 domestic AI chips, without a single Nvidia GPU.

A Food Delivery Company Pivoting to AI

Meituan officially launched LongCat-2.0 today (June 30) — but the real story actually began 22 hours earlier.

Without any prior announcement, Meituan quietly uploaded the model to OpenRouter (the world's largest AI model aggregation platform) under the anonymous codename Owl Alpha. The result? After silently running for 22 hours, it directly secured a Top 3 spot globally on OpenRouter's daily active rankings. Meituan's official Twitter account claimed it today, with even more detailed results:

• Hermes Agent: Global #1
• Claude Code: Global #2
• OpenClaw: Global #3

You read that right — these three are currently the most mainstream AI coding agent frameworks in the world. LongCat-2.0 directly swept the top three spots on developer usage rankings across these frameworks.

What makes this unusual: Meituan isn't an AI company — it's a food delivery company. A company that started by delivering meal boxes, spending 3 years in stealth to produce the strongest open-source model chosen by global developers through their own usage — that narrative alone defies common sense.

What's Actually Inside a Trillion-Parameter Model

A few hard numbers from LongCat-2.0 are worth remembering:

• 1.6 trillion total parameters / ~48 billion active (MoE architecture, dynamically activating 33B-56B)
• Native 1M context (million-level, not added later)
• 35 trillion tokens of training data
• 50,000 domestic AI chips (ASIC, not GPU)

On the architecture side, Meituan proposed three innovations of its own:

LSA (Latent Sparse Attention) — compresses long-text computation to linear growth, making 1M context no longer a game of throwing compute at the problem.

Zero-Compute Experts — dynamically adjusts activated parameters, activating only 33B-56B per token, wasting zero compute.

MOPD (Multi-expert Orchestration via Parallel Dispatch) — splits experts into Agent / Reasoning / Interaction groups, routing by task gating, so simple problems don't invoke the expensive group.

The most critical benchmark: 59.5 on SWE-bench Pro — a hard metric for evaluating real AI coding ability, where most Silicon Valley open-source flagships hover around this line. A quick note here: AIbase's earlier translation claiming "surpasses GPT-5.5 and Claude Opus 4.6" is an over-interpretation — Meituan's original wording was only "benchmarked against mainstream closed-source models."

The training-side numbers are even more worth discussing: 35 trillion tokens × 50,000 domestic chips × 3 years. Meituan revealed that through their self-developed deterministic operators and elastic recovery mechanisms, they reduced the average daily failure rate by over 70%, maintaining a stable daily throughput of 1T tokens. Three years were spent just to refine this training framework from "it runs" to "reliably producing trillion-parameter models."

The Price Is Genuinely Cheap

SiliconFlow launched it on Day 0 with this pricing (per 1M tokens):

For comparison, mainstream closed-source flagships (GPT-5, Claude Opus tier) typically sit at $3-15 for Input — LongCat-2.0's Input is roughly 1/4 to 1/20 of closed-source flagships. For Agent developers, this pricing means the same budget gets 4-20x the token volume, which is especially friendly for long-context tasks (codebase analysis, long-document summarization).

What This Means — Three Takeaways

Takeaway One: 50,000 domestic chips ≠ a stopgap; the full pipeline is genuinely working.

Many people's impression of domestic AI chips was "fine for inference, but training still needs Nvidia." LongCat-2.0 is the first trillion-parameter model with full-cycle training + inference running entirely on domestic ASIC clusters. This means the entire suite of industrial-grade challenges — from operator adaptation and communication optimization to distributed stability — has been genuinely solved with domestic chips + domestic frameworks, not at a demo level.

Takeaway Two: The "cost-performance war" among open-source flagships is still accelerating.

After DeepSeek brought down closed-source model prices, domestic open-source flagships have been chasing ever since. LongCat-2.0's pricing is essentially "one more cut" — and this time it's not subsidy-driven, but real cost savings from full autonomy across hardware and the training stack. If the closed-source camp doesn't follow, they either lose the long-context market or are forced to cut prices — a lose-lose.

Takeaway Three: Meituan's ambition isn't food delivery.

Spending 3 years building this model, then directly hitting global top 3 across Hermes / Claude Code / OpenClaw agent frameworks — Meituan is using a "model + toolchain" bundled strategy to capture developer mindshare. It's the same logic as ByteDance capturing users with TikTok's algorithm: first lock in global developers with an open-source model, then monetize through toolchains and applications.

Personal Take

LongCat-2.0 may not be the strongest model right now (OpenRouter Top 3 is by usage volume, not absolute capability), but it proves one thing: domestic AI chips can already support industrial-grade training of trillion-parameter models. That matters more than any single model's benchmark score — it directly overturns the industry perception that "domestic chips can only handle inference."

Over the next 6 months, the thing to watch is whether domestic AI chip makers like Huawei Ascend, Cambricon, and Hygon will ride LongCat-2.0's momentum to push "training card + inference card" full-stack solutions into the enterprise market. If they do, Nvidia's last stronghold in the Chinese market (trillion-parameter model training) will be formally breached.

What Do You Think

LongCat-2.0's "anonymous trial + sudden claim" launch strategy — do you see it as low-key or calculated?

— And the sharper question: Do you think Meituan built this model genuinely for AI, or to give its food delivery business a new "tech company" story?

Sound off in the comments.