Qwen2.5 32B Instruct: how much VRAM does it need?

Qwen2.5 32B Instruct has 32.76B parameters in 64 layers. It uses grouped-query attention: 40 query heads share 8 key-value heads, so its KV cache is 5× smaller than full multi-head attention would need. It takes up to 32,768 tokens of context; one user at that length needs 8.00 GiB of KV cache at FP16 on top of the weights.

BF16 · 8k context · one user

65.64 GiB

61.02 GiB of weights, 2.00 GiB of KV cache and 2.62 GiB of vLLM's overhead. At INT4 the same workload needs 22.96 GiB.

64 layers
hidden 5120
40/8 heads
head_dim 128
32k max context

Qwen2.5 32B Instruct memory by precision and context

Precision	8k context	32k context
BF16	65.64 GiBKV cache 2.00 GiB	71.64 GiBKV cache 8.00 GiB
FP8	36.58 GiBKV cache 2.00 GiB	42.58 GiBKV cache 8.00 GiB
INT4	22.96 GiBKV cache 2.00 GiB	28.96 GiBKV cache 8.00 GiB

One user, FP16 KV cache and vLLM's overhead on one GPU at its defaults for a card under 70 GB. Quantizing shrinks the weights, not the KV cache, so the saving narrows as the context grows.

Which single GPUs hold it

BF16 (65.64 GiB): A100 80GB, H100 SXM 80GB, H100 PCIe 80GB, H100 NVL 94GB
FP8 (36.58 GiB): NVIDIA L40S, RTX 6000 Ada, RTX A6000, AMD Ryzen AI Max+ 395 (64GB)
INT4 (22.96 GiB): RTX 5090, Apple M4 Pro, A100 40GB, NVIDIA L40S

At 8k context for one user, the smallest cards first. A discrete card is budgeted at vLLM's default 92% of its memory, a unified-memory device at the share its GPU can address, and each carries vLLM's overhead at its own defaults.

How many users fit

GPU	BF16	FP8	INT4
RTX 4090	does not fit	does not fit	does not fit
RTX 5090	does not fit	does not fit	3
RTX 6000 Ada	does not fit	4	11
H100 SXM 80GB	4	18	25
RTX PRO 6000 Blackwell	11	26	32
AMD Ryzen AI Max+ 395 (128GB)	12	26	33
H200 141GB	32	46	53
B200 180GB	48	63	69

Users at 8k context each, on one card. ✗: the weights leave no room for even one.

Fine-tuning Qwen2.5 32B Instruct

Method	Memory	Smallest GPU
Full fine-tune	490.4 GiB	several GPUs, sharded
LoRA (rank 16, attention)	63.71 GiB	RTX PRO 5000 Blackwell (72GB)
QLoRA (rank 16, attention)	18.43 GiB	NVIDIA L4

Sequence length 2,048, batch size 1, gradient checkpointing on, Adam. Full fine-tuning keeps 16 bytes per parameter before activations; QLoRA stores the frozen base in 4-bit NF4.

Where the numbers come from

Weights are parameters × bytes per parameter. The KV cache is 2 × layers × key-value heads × head dimension × bytes × context × users. For Qwen2.5 32B Instruct that is 256.0 KiB per token at FP16. These are capacity-planning figures: the serving engine — vLLM, TGI, llama.cpp — and its paged-attention efficiency still move the real number.

More on KV cache size and on GQA, MQA and MLA.