What GPU do you need to run LongCat-2.5?
Confirmed setup not yet documented. MIT open weights; 3,561.97 GB BF16 download, not memory. Cluster-scale NPU recipe only, no PC/Mac. No Q4, no fit or speed.
1779.75B parameters, 48B active (mixture-of-experts) (Hugging Face, accessed 2026-11-09). No Q4 GGUF is sourced, so this page does not estimate a fit or a speed. The weight download is 3,561,974,336,256 bytes (3,561.97 GB) across 210 safetensors shards (source). That is the size of the files, not the memory a runtime keeps resident. The 210 files are 208 model shards (3,553,858,076,424 bytes) and two multimodal model_mm files (8,116,259,832 bytes). The repo's used storage is 3,561,989,945,589 bytes. Commit 6fc4844cc9455bc45b84871cbc9137b3286f0fc2. No context estimate: LongCat-2.5's MLA and related cache layout has no verified llama.cpp KV formula yet (details).
48B active is the model card's count of parameters used for one token. It is not the memory the checkpoint occupies, and the weight download is not that memory either. The card claims 1.6 trillion total parameters, and Hugging Face's safetensors metadata sums to 1,779,752,648,320 (1,779.75B on this page). Those are two published figures. This page lists both and does not guess at the difference. config.json sets a 524,288-token context, which is the window shown here. Some evaluation footnotes and the NPU launch script use about 1 million tokens, and this page does not treat that as a confirmed window. The official W4A4 checkpoint is 1,336.01 GB of safetensors for NPU serving. It is not a GGUF and not a file the calculator can score. The 194-NPU Atlas A2 layout in Meituan's serving guide is one documented recipe, not a minimum. We have not run LongCat-2.5, and no speed or rental price is listed.
What we checked
Checked Nov 9, 2026. Official is the maker or the weight files. A vendor claim is that project's own documentation. A community claim is someone else's upload. Our check is a file or search we read. Unknown means we have no source, so it is not estimated. None of these rows is a speed we measured.
| Kind | What | Detail | Source |
|---|---|---|---|
| Official | Model card | Meituan's card describes LongCat-2.5 as a native multimodal mixture-of-experts model and claims 1.6 trillion total parameters with about 48 billion activated per token. It lists the license as MIT. Its deployment section documents NPU serving through SGLang-FluentLLM and nothing else: no NVIDIA GPU, llama.cpp, MLX, or vLLM recipe. Hugging Face's safetensors metadata totals 1,779,752,648,320 parameters (1,779.75B on this page; the Hugging Face page rounds it to 1.8T). The card's 1.6T and the metadata's 1.78T are two published figures, and this page does not guess why they differ. The 48B figure is not resident memory. | Model card, checked Nov 9, 2026 |
| Official | Weight files | 210 safetensors files sum to 3,561,974,336,256 bytes (3,561.97 GB): 208 complete model shards (model_*-of-00208) at 3,553,858,076,424 bytes and two model_mm files at 8,116,259,832 bytes. The repo's used storage is 3,561,989,945,589 bytes, and the full file tree is 3,562,006,190,340 bytes. The metadata splits the parameters into 1,778,523,785,600 BF16 and 1,228,862,720 F32. The repo was created 2026-11-04 and last modified 2026-11-09 (commit 6fc4844cc9455bc45b84871cbc9137b3286f0fc2), the day of the public announcement. We read the sizes from the Hugging Face API and did not download the files. They are download sizes. They are not GPU memory, unified memory, or a GGUF. | Hugging Face API, checked Nov 9, 2026 |
| Official | Architecture config | config.json sets model_type pro_v and architecture LongcatCausalLM: 38 layers with MLA attention, 768 routed experts with 12 selected per token (moe_topk), and 128 identity (zero-cost) experts. max_position_embeddings is 524,288, the context shown on this page. Some evaluation footnotes and the NPU launch script use about 1 million tokens. This page does not treat that as the model's confirmed window. | config.json, checked Nov 9, 2026 |
| Official | License | The LICENSE file is the standard MIT license, Copyright (c) 2026 Meituan. It allows use, copying, modification, distribution, and sale, as long as the copyright and permission notice stay with the copies. This is a summary of the license, not legal advice. | LICENSE, checked Nov 9, 2026 |
| Official | W4A4 weight files | Meituan also publishes LongCat-2.5-W4A4. Its 210 safetensors files sum to 1,336,010,166,464 bytes (1,336.01 GB), with all 208 main shards present. The full file tree is 1,336,050,649,636 bytes and used storage is 1,336,042,442,734 bytes; the repo was last modified 2026-11-09. This is the checkpoint the NPU guide serves. It is not a GGUF and not a verified fit on any consumer machine, and this site does not use its size as a Q4 file. | W4A4 repo, checked Nov 9, 2026 |
| Official | Hosted chat | Meituan's LongCat site offers a chat interface. That is Meituan's service, not a local setup, and this page lists no price for it. | longcat.ai, checked Nov 9, 2026 |
| Vendor claim | SGLang-FluentLLM NPU guide | The model card points to SGLang-FluentLLM. Its longcat-2.5-npu branch has the documented LongCat-2.5 recipe: the W4A4 checkpoint on Atlas A2 with an Encoder-Prefill-Decode split across 194 NPUs (2 encoder, 64 prefill, 128 decode). That is one documented cluster recipe. It is not a minimum hardware requirement and not a memory floor. The project's main README mentions NVIDIA H800 and H20 testing, but its Quick Start covers LongCat-Flash-Lite, and no LongCat-2.5 GPU recipe is documented. We have not run either. | NPU guide, checked Nov 9, 2026 |
| Our check | Loader code and encoder weights | config.json's auto_map names configuration_longcat_oe.py and modeling_longcat_oe.py, and the repo is tagged custom_code, but neither the Hugging Face nor the ModelScope file tree has any .py file. The NPU guide loads weights from a LongCat-2.5-Encoder path, and we found no public Hugging Face or ModelScope repo under that name. Meituan has no LongCat-2.5 model repo on GitHub; the serving code is in SGLang-FluentLLM. | Hugging Face file tree, checked Nov 9, 2026 |
| Our check | No GGUF, and no mlx-community port | Hugging Face searches for LongCat-2.5 filtered to GGUF and to MLX returned no models, and mlx-community has no LongCat-2.5 repo. Ports of the older LongCat-Flash models are different models and are not listed here. | Hugging Face GGUF search, checked Nov 9, 2026 |
| Our check | llama.cpp | llama.cpp master has no longcat or pro_v architecture. Open pull requests add LongCat-Flash and LongCat-Flash-Lite, which are different models, and none of them is merged. | llama.cpp architectures, checked Nov 9, 2026 |
| Our check | mlx-lm | mlx-lm's model folder has longcat_flash.py and longcat_flash_ngram.py, and no file for LongCat-2.5's pro_v architecture. | mlx-lm models, checked Nov 9, 2026 |
| Unknown | Not documented | Unknown, and not estimated here: how much memory a runtime keeps resident, KV-cache bytes per token, any measured tokens per second, a LongCat-2.5 GPU recipe, a tested PC or Mac in our catalog, and a rental price. | Model card does not give this, checked Nov 9, 2026 |
Other places to get it
Checked Nov 14, 2026. We list only copies we opened and checked. Check file hashes for copies from other uploaders.
- ModelScope: Lab's own account