LongCat-2.0: Meituan’s 1.6T Open-Source Coding Model Trained on Chinese Chips

LongCat-2.0 is the model the AI world woke up to on June 30, 2026. Meituan, the Beijing-based on-demand services giant, open-sourced a 1.6-trillion-parameter coding model built start to finish on domestic Chinese chips. Moreover, it is already topping OpenRouter for agentic coding and edges GPT-5.5 on a closely watched benchmark. As a result, it has forced a quick rethink of what open, “made in China” AI can actually deliver.

LongCat-2.0 is a 1.6-trillion-parameter, open-source large language model built specifically for agentic coding. Instead of firing every parameter at once, it uses a Mixture-of-Experts design. That design activates only 33 to 56 billion parameters per token, roughly 48 billion on average. Because of that, it can carry a native one-million-token context while keeping inference costs sane.

The model did not appear out of nowhere. Meituan shipped LongCat-Flash, a 560-billion-parameter model, back in September 2025, then a multimodal LongCat-Next in March 2026. This release lands under the permissive MIT license, with the weights published on Hugging Face under the meituan-longcat organization. In short, anyone can download it, study it and build on it.

Under the hood, the architecture does the heavy lifting. The model leans on LongCat Sparse Attention, an in-house evolution of the sparse-attention approach DeepSeek popularized, to keep that million-token window affordable. Meanwhile, a multi-expert post-training system splits the work into dedicated Agent, Reasoning and Interaction expert groups, so the model can plan, think and chat without stepping on its own toes.

The training run was enormous. Meituan fed it more than 30 trillion tokens spanning Chinese, English, other languages and code. Crucially, the result plugs straight into the tools developers already use: it integrates natively with Claude Code, OpenClaw and Hermes for repository-level edits and automated task execution. Therefore it behaves less like a chatbot and more like a coworker that can act across a whole codebase.

On Meituan’s own numbers, LongCat-2.0 posts a 59.5 on SWE-bench Pro, nudging past GPT-5.5’s 58.6. It also records 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual and 73.2 on the FORTE corporate-workflow simulator. Taken together, the company argues overall performance is comparable to Google’s Gemini 3.1 Pro. That puts it among the most closely watched of the Latest AI Models 2026.

However, a caveat belongs up front: these are self-reported figures. Independent reproduction on third-party leaderboards was still pending at launch. So treat the scores as a strong early signal, not a settled verdict.

The benchmark that may matter most is not on any leaderboard. Meituan says this is the industry’s first trillion-parameter model to finish both training and inference on a 50,000-card domestic compute cluster, built from large-scale pods of AI ASIC accelerators. Notably, its use of the Huawei Collective Communication Library points to Huawei silicon, with observers betting on Ascend 910C chips.

That detail carries real weight. DeepSeek’s earlier V4-pro leaned on domestic hardware only for inference, while the heavy pre-training still ran elsewhere. By doing the whole job at home, Meituan signals that frontier-scale training no longer depends on Nvidia or AMD. In a year defined by US export controls, that is a genuinely strategic message.

Context matters, so here is the honest scoreboard. LongCat-2.0 beats GPT-5.5 on SWE-bench Pro and slips past Gemini 3.1 Pro on both SWE-bench Pro (59.5 to 54.2) and Terminal-Bench (70.8 to 70.7). Against Anthropic’s Opus 4.8, though, it trails across every code-agent metric, including a 78.9 Terminal-Bench and a 69.2 SWE-bench Pro.

The fair summary is near-frontier, not frontier-beating. It is not the best coding model on Earth. Instead, it is an open, free-to-download model that lands within striking distance of the closed leaders — and it does so on hardware Washington tried to keep out of reach.

Getting hands-on is the easy part; running it locally is not. The fastest route is the hosted API at longcat.chat, while the open weights arrive on Hugging Face for teams that want to self-host. Be warned, though: even at aggressive 2-bit quantization the model needs 400GB-plus of storage, so as one early write-up put it, if you have to ask whether your GPU can run it, it cannot.

For most developers, that means calling it through an API or a hosted provider and slotting it into an agentic workflow next to the other best AI coding tools. Used that way, its million-token context and native Claude Code support make it a serious option for large, messy repositories.

Step back, and the release is bigger than any single leaderboard. A trillion-parameter, MIT-licensed model trained on domestic chips hands every researcher and startup a frontier-class base they can actually own. Consequently, the gap between closed labs and open weights keeps shrinking, month after month.

It also extends a clear Chinese trend. From DeepSeek to the GLM family, open-weights models keep undercutting Western incumbents on price and, increasingly, on coding. Meituan just raised the stakes by proving the entire stack — chips included — can be home-grown.

If this release has you tracking the open-weights race, read our breakdown of GLM 5.2, another Chinese model undercutting GPT-5.5 on coding. And when you are ready to put a model to work, compare the best vibe coding tools that pair perfectly with an agentic LLM.

*Sources: VentureBeat, South China Morning Post, Crypto Briefing, best-ai.news.*