Unconfirmed: a post circulating among model watchers describes a large Chinese model released with unusually little supporting material. Automatica has not verified the specifications, the pricing or the absence of benchmarks, and the account is the only source for them.
The account TeksEdge wrote that "Meituan dropped LongCat-2.5-Preview" and listed the specifications it had seen: "1.6 TRILLION parameters", "~48B active/token", a "1 MILLION token context", native multimodality and a design "built for long-running agents". It quoted prices of "$0.30 / 1M input tokens", "$0.006 / 1M cached input" and "$1.20 / 1M output", and said the model can already be used from several agent tools.
The post's own framing is the interesting part. "But here's the strange part... where are the benchmarks?" it asks, before answering: "No benchmark table. No LongCat-2.5 Hugging Face weights I can find. No GGUF." Its opening line — "Okay this is a weird big LLM launch. Why so low-key??" — is the claim worth testing, more than any individual number.
A model of that size released without published evaluations or downloadable weights is not unheard of; a preview endpoint with pricing attached and nothing to reproduce is a familiar pattern for vendors testing demand. It does mean the headline figures rest entirely on the provider's word, and in this case on one reader's account of the provider's word.
What would settle it: a model card, a benchmark table, weights or a paper from Meituan, or independent evaluations from people who have paid for the endpoint and measured it. Meituan has published none of those as far as this post's author could find, and Automatica has not found them either.
Readers should treat every number above as a claim about a claim. We are reporting that a launch of this scale is being discussed on the basis of a single account's specification list, which is itself the story.