Meituan putting a 1.6T-parameter model under MIT, trained on domestic accelerators, strengthens my open-weights bull case and adds one…

Meituan putting a 1.6T-parameter model under MIT, trained on domestic accelerators, strengthens my open-weights bull case and adds one more data point to my mixed read on Nvidia.

• Meituan's Nov 10 post says LongCat-2.5 is open-sourced under MIT, with commercial deployment and further training allowed without extra permission. It's 1.6T total parameters, about 48B active per token, roughly a 1.8TB checkpoint.
• Meituan reports Terminal Bench v2.1 at 88.4 versus 70.8 for the prior model, and DeepSWE v1.1 at 64.6 versus 9.7. Those are company numbers until someone independent reproduces them.
• Meituan says training and fine-tuning used domestic accelerators. The closest comparison I'd make is to my own Nvidia flip condition: a cheaper path to frontier-scale training is the kind of thing I watch for.
• Hosted pricing is $0.30 per million input tokens and $1.20 per million output. At that cost, the incumbents' pricing power keeps shrinking.

On the stock, the -7.34% MPNGY move in the Nov 9 session lines up with Morgan Stanley's downgrade on slowing e-commerce growth, a day before Meituan's model announcement, so I'd keep the two stories apart. My read: cheap MIT weights are hard to stop, and I'll take the benchmarks seriously once outside testers confirm them. Could be wrong on how fast that happens.

Sources:
https://aiweekly.co/alerts/meituan-open-sources-longcat-25-16t-moe-under-mit-license
https://news.futunn.com/en/flash/1003981562/meituan-open-sources-longcat-2-5-model-1-6-trillion

Not financial advice. For informational and educational purposes only.