Meituan Releases LongCat-2.0: A 1.6-Trillion-Parameter Coding Model Trained on 50,000 Domestic Chips

China's Meituan released the open-source coding model LongCat-2.0 on June 30, 2026. LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts (MoE) model trained from scratch on a 50,000-chip cluster of domestic processors, with no U.S. accelerators. It is specialized for agentic coding, supports a 1-million-token context window, and is published under the meituan-longcat organization on Hugging Face. ASAP summarizes the facts in a direct-answer format based on primary reporting from Crypto Briefing and Reuters.

Hardware Independence Is the Real Headline

Meituan said on June 30, 2026, that LongCat-2.0 was trained from scratch on a cluster of 50,000 domestic chips. According to Crypto Briefing, the training used no U.S. export-restricted accelerators such as Nvidia's A100 and H100 or AMD's MI300X.

The center of gravity in this news is the training infrastructure, not the model's scores. Chinese labs have run inference and fine-tuning on domestic silicon before, but publicly stating that the full from-scratch pretraining was completed on domestic hardware alone is rare. With export controls tightening, that declaration itself functions almost like a bargaining chip.

How to Read the 1.6 Trillion Parameters

LongCat-2.0 is a 1.6-trillion-parameter Mixture-of-Experts model that dynamically activates 33 to 56 billion parameters per token. It accepts inputs of up to 1 million tokens to work with ultra-long documents and large codebases.

Item	Value
Release date	June 30, 2026
Total parameters	1.6 trillion (MoE)
Active per token	33–56 billion
Context window	1 million tokens

Do not misread the number. The 1.6 trillion is total stored capacity, while only about 3 percent of it, the 33 to 56 billion, actually fires on any single inference pass. In other words, the footprint is enormous but the runtime cost is held down to a mid-sized model. That is also why brute scale could compensate even if each domestic chip trails the top-tier GPUs on raw per-unit performance.

Benchmark Scores and Their Confidence Interval

Meituan reported that LongCat-2.0 scored 59.5 on the SWE-bench Pro coding benchmark and 70.8 on Terminal-Bench. Meituan claimed the model matched or exceeded leading proprietary models such as Google's Gemini, OpenAI's GPT-5.5, and Anthropic's Claude Opus on some coding and agent benchmarks. The benchmark figures are Meituan's own reported numbers, and independent verification was not yet available at the time of release.

It is worth stepping back here. Self-reported scores are sensitive to which benchmarks are chosen and how they are measured, and the comparison to proprietary models omits exactly which versions were pitted against it. Being open source actually helps on this front. Because the weights are public, the community will soon rerun the same benchmarks, and until those reproductions land, 59.5 and 70.8 are best treated as reference lines rather than settled results.

What It Means for Korean Developers and Companies

Meituan made LongCat-2.0 downloadable as open source under the meituan-longcat organization on Hugging Face. Earlier, players such as DeepSeek similarly used open-source releases to broaden their developer base amid the export-control environment.

For Korean practitioners the practical upside is clear. The weights can be pulled and hosted on internal servers, so code never leaves for an external API, and the 1-million-token context window is immediately useful for analyzing large legacy codebases. That said, the license terms and any dependence on domestic-chip-optimized kernels are things to verify before adoption. Following DeepSeek, this release widens the menu of open-weight coding models by another notch, giving teams a concrete reason to reconsider sole reliance on a single commercial API.

---

Sources: Crypto Briefing · AOL/Reuters · Hugging Face (reported June 30, 2026)