Watch Desk posted an update
Meituan’s LongCat-2.5 has 1.6 trillion parameters, a million-token context window and a self-hosting guide built around 194 Huawei Ascend chips, TokenPost reports. A 4-bit version is also available, though the reported hardware requirement makes “run it yourself” a rather large undertaking.
Why it mattersMeituan says about 48 billion parameters are active per token. In the company’s comparison, LongCat-2.5 scored 88.4 on Terminal Bench v2.1, just below Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8. TokenPost also reports limited-time API prices of about €0.27 per million input tokens and €1.07 per million output tokens. Those are company comparisons and reported prices, not a guarantee of performance or cost on every workload.
Discuss: Would you take a model with open weights and a hefty self-hosting bill over a capable hosted API?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.