Meituan (LongCat), China · Released November 4, 2026New
LongCat-2.5 has 1779.8B parameters (48B active per token). In BF16 its weights alone take 3058 GiB; quantized to 4-bit, about 922 GiB. Each 32K-token request adds 2.82 GiB of KV cache. Cheapest way to run it today: 6 × AMD MI300X (192 GB) at $14.34 per hour. It reads text, images and video, calls tools and has a thinking mode you can switch off.
Weights plus the KV cache of one request (BF16 cache), in GiB. Add 1 to 3 GiB for the inference engine itself.
Weights format
Weights only
+ 8K context
+ 32K context
+ 128K context
+ 512K context
BF16 / FP16
3058
3059
3061
3069
3103
FP8 / INT8
1531
1532
1534
1543
1577
4-bit (GGUF Q4_K_M)
922
922
924
933
967
How many GPUs to run LongCat-2.5
Fewest GPUs that fit the weights plus one 8K-token request. Data-center GPUs use vLLM (90% of memory usable), the others llama.cpp. Several GPUs split the model with tensor or pipeline parallelism.
GPU
BF16 / FP16
8-bit
4-bit
GeForce RTX 4090 24 GB
> 128
96 (several machines)
48 (several machines)
GeForce RTX 5090 32 GB
128 (several machines)
64 (several machines)
32 (several machines)
RTX PRO 6000 Blackwell 96 GB
32 (several machines)
16 (several machines)
16 (several machines)
Mac M5 Max 128 GB 96 GB usable
32 (several machines)
16 (several machines)
16 (several machines)
Ryzen AI Max+ 395 128 GB 96 GB usable
32 (several machines)
16 (several machines)
16 (several machines)
DGX Spark (GB10) 128 GB
24 (several machines)
16 (several machines)
7
NVIDIA H100 80 GB
48 (several servers)
24 (several servers)
16 (several servers)
NVIDIA H200 141 GB
32 (several servers)
16 (several servers)
8
NVIDIA B200 (HGX) 180 GB
24 (several servers)
16 (several servers)
6
8-bit: FP8 with vLLM (INT8 on GPUs without FP8), Q8_0 with llama.cpp. 4-bit: AWQ / GPTQ with vLLM, Q4_K_M with llama.cpp, both counted at the Q4_K_M size.
Cheapest way to run LongCat-2.5 today
For each GPU: the fewest cards that fit LongCat-2.5 with one 8K-token request, the most faithful weight format at that count, and the cheapest on-demand price (2026-11-22).
On-demand prices from AWS, Google Cloud, Azure, RunPod, Vast.ai, Oracle Cloud, OVHcloud, Scaleway, Verda, Akash and UpCloud, checked every hour. Serving many users needs more KV cache, so more memory: size it in the calculator. Some links are affiliate links: StudioTV may earn a commission, at no extra cost to you.
Reasoning. Thinking mode you can switch on or off. With vLLM or SGLang, set it per request: "chat_template_kwargs": {"enable_thinking": false} turns it off, true on. Reasoning tokens take context and KV cache like any other token: leave room for them.
KV cache: how memory grows with context
Architecture: MoE 768 experts + 128 zero experts · 2 MLA attentions per layer: 76 caches (512 + 64 latent), sparse-attention indexer on 38 · 135B n-gram embeddings · image and video input. Its cache stays compact because MLA stores one compressed latent per token and layer instead of full keys and values. Only 48B parameters are active per token, but all 1780B must sit in memory: VRAM depends on the total, speed on the active part.
Context per request
BF16 cache
FP8 cache
8K tokens
0.71 GiB
0.42 GiB
32K tokens
2.82 GiB
1.67 GiB
128K tokens
11.3 GiB
6.70 GiB
512K tokens
45.2 GiB
26.8 GiB
LongCat-2.5 VRAM FAQ
How much VRAM does LongCat-2.5 need?
In BF16 the weights alone take 3058 GiB (3283 GB). In FP8 that is 1531 GiB, and about 922 GiB with 4-bit quantization (Q4_K_M). Each request then adds KV cache: 0.71 GiB at 8K tokens and 2.82 GiB at 32K (BF16 cache).
Can LongCat-2.5 run on a single RTX 4090 (24 GB)?
No. Even in 4-bit it needs 48 RTX 4090s, spread over several machines. In 4-bit it needs 16 H100 GPUs, across several servers.
Can LongCat-2.5 run on an RTX 5090 (32 GB)?
No. Even in 4-bit it needs 32 RTX 5090s, spread over several machines.
Can LongCat-2.5 run on a Mac?
No single Mac holds it, even in 4-bit: it needs 922 GiB with an 8K-token context, and the largest Mac (512 GB) gives its GPU about 384 GB.
Can LongCat-2.5 run on a DGX Spark or a Ryzen AI Max+ 395 (128 GB)?
No. Even in 4-bit it needs 7 DGX Spark units or 16 Ryzen AI Max+ 395 machines (128 GB), linked over the network.
What is the cheapest way to run LongCat-2.5?
On 2026-11-22, the cheapest on-demand setup is 6 × AMD MI300X (192 GB) with 4-bit weights on RunPod, at $14.34 per hour (about $10,468.20 per month). Next: 4 × AMD MI355X / MI350X (288 GB) with 4-bit weights on RunPod, at $21.96 per hour. Sized for one 8K-token request; prices are checked every hour.
How many H100 GPUs does LongCat-2.5 need?
With one request and an 8K-token context: 48 in BF16, 24 in FP8 and 16 in 4-bit (vLLM, 90% of memory usable). Above 8 GPUs, that means several servers. Serving many users at once needs more memory for their KV caches: the calculator sizes that for you.
Can LongCat-2.5 read images?
Yes. LongCat-2.5 reads text, images and video. Images add tokens to the context, so they also add KV cache.
Does LongCat-2.5 support tool calling?
Yes, LongCat-2.5 supports tool (function) calling: it can ask your application to run a function and use the result.
Does LongCat-2.5 have a thinking mode?
Thinking mode you can switch on or off. With vLLM or SGLang, set it per request with "chat_template_kwargs": {"enable_thinking": false} (or true). Reasoning tokens count toward the context and the KV cache.
LongCat-2.5 VRAM badge
For a model card or a README: the 4-bit size and the smallest GPU it fits on, linked to this page. Data as JSON: /api/models/longcat-2-5.json.
Estimates, not guarantees: computed from the official config.json with the same engine as the LLM VRAM Calculator. Real usage depends on your engine version and settings.