June 29, 2026
From takeout to takeover?
LongCat-2.0, a large-scale MoE model with 1.6T total and 48B Active
A giant new AI just dropped, and the internet is already asking who built it and what it really is
TLDR: LongCat-2.0 is a huge new open AI model claiming stronger coding skills and training on alternative chips instead of the usual Nvidia setup. Commenters immediately turned it into a drama thread, arguing over whether it’s truly original, what hardware was really used, and why a food delivery company is suddenly in the AI big leagues.
LongCat-2.0 arrived with a massive flex: its creators say it’s an open-source AI trained at enormous scale, with a huge memory window and stronger coding skills, all running on non-Nvidia hardware. In plain English, this is a very big chatbot brain that can handle long documents and complicated tasks. But in the comments, people were way less impressed by the chest-thumping and way more interested in the plot twists behind it.
The loudest reaction? "Wait, is this basically DeepSeek in a different outfit?" One commenter politely but very pointedly wondered whether LongCat-2.0 is truly its own thing or mostly a fine-tuned version of an existing Chinese model. That kicked off the classic open-model drama: breakthrough, remix, or just smart repackaging? Another crowd favorite said the real story isn’t the model at all — it’s the claim that the training happened on huge AI chip clusters outside the usual Nvidia universe, with some readers immediately speculating about Huawei. That turned the launch into a hardware soap opera.
Then came the comedy. One commenter stress-tested the model with a bizarre nuclear-fuel question, basically treating the release thread like a live talent show. Another begged for the one thing ordinary users actually want: can it run on llama.cpp, and how fast on normal gear? And perhaps the most memeable reveal of all: people were stunned this model appears tied to Meituan, a food delivery company. Nothing says 2026 tech chaos like ordering noodles from the same corporate family that may also want to automate your coding workflow.
Key Points
- •LongCat-2.0 is an open-source mixture-of-experts language model with 1.6 trillion total parameters and about 48 billion active parameters per token.
- •The article says both training and deployment were carried out entirely on AI ASIC superpods.
- •Pretraining reportedly spanned millions of accelerator-days and used more than 35 trillion tokens without rollbacks or irrecoverable loss spikes.
- •The model adds LongCat Sparse Attention and was trained on hundreds of billions of tokens of 1M-context data to improve long-horizon performance.
- •LongCat-2.0 is described as integrated with Claude Code, OpenClaw, and Hermes for coding, repository editing, automated task execution, and agentic workflows.