Meituan Open-Sources LongCat-2.0, First Trillion-Parameter Model Trained Entirely on Chinese Chips

Chinese conglomerate Meituan, best known for its food delivery app, has released the LongCat-2.0 language model under an open license, with 1.6 trillion parameters. What matters most isn't the sheer scale but how the model came to be: the entire training and inference process ran on a cluster of 50,000 Chinese ASIC chips, without a single Nvidia GPU involved.

Two Months in Hiding

Before Meituan officially unveiled the model on June 30, LongCat-2.0 had already been running anonymously on the OpenRouter platform for two months under the pseudonym Owl Alpha. During that time, it climbed to first place in the Hermes Agent workspace, second place among models used in Claude Code, and third place in OpenClaw deployments, ranked by monthly query volume. Only the launch under its real name revealed who was behind the results.

Architecture and Scale

LongCat-2.0 is a mixture-of-experts model with 1.6 trillion total parameters, of which between 33 and 56 billion are activated at once depending on query complexity. The model natively supports context windows of up to one million tokens thanks to the LongCat Sparse Attention mechanism, optimized for long sequences. Training covered more than 35 trillion tokens of data across multiple languages and code, and Meituan says the entire process ran without major crashes or lost progress requiring a rollback to earlier checkpoints.

On the SWE-bench Pro benchmark, the model scored 59.5 points, ahead of GPT-5.5's 58.6, though it trailed Anthropic's top models. On the FORTE benchmark, LongCat-2.0 scored 73.2 points, tying Claude Opus 4.6 and trailing GPT-5.5, which reached 77.8 points. The results place the model among near-frontier systems, despite being built without any Nvidia hardware.

Pricing Far Below Competitors

LongCat-2.0's standard API pricing is $0.75 per million input tokens and $2.95 per million output tokens, dropping to $0.30 and $1.20 during the promotional period. By comparison, GPT-5.5 costs $5 and $30 per million tokens, while Claude Sonnet 5 costs $2 and $10 respectively. Context cache reads are free, and Meituan also offers token bundles priced at around $60 per billion.

What It Means for China's Tech Independence

LongCat-2.0 differs fundamentally from the earlier DeepSeek V4-Pro, which was trained on Nvidia chips with only inference running on Chinese Huawei hardware. Here, the entire chain, from training to serving queries, ran on domestic hardware, something Meituan and industry commentators call the first case of its kind for a model of this scale. Training is far more computationally demanding than inference alone, so proving it can be done without American chips carries a strategic weight beyond the model's specs.

The model's weights are expected to land on Hugging Face and GitHub under a progressively wider release, though for now they are marked as coming soon, with no firm date. Programmatic access is already available through an API compatible with OpenAI and Anthropic formats, with integrations including Hermes, Claude Code, and OpenClaw.

For Polish companies and developers using models via API, this marks another low-cost alternative for agentic and coding tasks, though self-hosting on private infrastructure isn't available yet. The launch also shows that US export restrictions on AI chips to China haven't stopped Chinese firms from building models competitive with the Western frontier, only accelerated investment in domestic hardware.

Sources: Meituan Open Sources LongCat-2.0 (venturebeat.com), China Debuts Biggest AI Model Trained on Local Chips (scmp.com), LongCat-2.0: The Stealth AI Model That Was Quietly Topping OpenRouter All Along (decrypt.co).