LongCat-2.5
Licence: MIT — free for any use, including commercial. Just keep the licence notice.
Meituan’s native multimodal successor to LongCat-2.0 for long-horizon agentic work: 1.6T total parameters, ~48B active per token, weights released on Hugging Face under MIT (repository created Nov 4, 2026; shipped BF16 files total about 3,562GB, so even a 4-bit build at roughly 800GB is a multi-node job). Watch-out: the model card’s only documented deployment route is NPU platforms via SGLang-FluentLLM, with a W4A4 scheme built for domestic accelerators — no GPU serving recipe is published on the card, so GPU self-hosting is unverified for now. Maker-reported, in-house scores on the card include Terminal-Bench v2.1 88.4 (LongCat-2.0: 70.8).
What it needs
How we work this out: memory = the model's total parameters × bytes per weight — 2 bytes at full precision (FP16), 1 byte at 8-bit, half a byte at 4-bit. Weights only — add room for the conversation memory (KV cache) and the serving software. A GPU counts as running a build when its memory covers the weights plus about 20% headroom. Real figures move with context length and software.
GPUs that can run it
Ranked by how well the model runs, then by price.
No single GPU can run LongCat-2.5 — even its 4-bit build needs 800GB, plus room to work, and no single card holds that. Needs at least 4× B300 by memory at 4-bit — a memory fact, not a performance promise. The prices board shows the biggest cards and what each is short by, with the single-GPU and node prices a setup like this is built from.
The minimum setup, by memory
Needs at least 4× B300 by memory at 4-bit — a memory fact, not a performance promise.
No single card runs LongCat-2.5 alone at this size — it is a multi-GPU job. The board lists the cards, nodes and racks a setup is built from, with node offers shown by their total price.
See the GPUs for LongCat-2.5 →
Checked Nov 15, 2026
Select a GPU See these GPUs on the prices board, ranked for this model.