Qwen2.5 14B Instruct: how much VRAM does it need?

Qwen2.5 14B Instruct has 14.77B parameters in 48 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 6.00 GiB of KV cache at FP16 on top of the weights.

BF16 · 8k context · one user

31.48 GiB

27.51 GiB of weights, 1.50 GiB of KV cache and 2.46 GiB of vLLM's overhead. At INT4 the same workload needs 13.40 GiB.

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

Qwen2.5 14B Instruct memory by precision and context

Precision	8k context	32k context
BF16	31.48 GiBKV cache 1.50 GiB	35.98 GiBKV cache 6.00 GiB
FP8	19.17 GiBKV cache 1.50 GiB	23.67 GiBKV cache 6.00 GiB
INT4	13.40 GiBKV cache 1.50 GiB	17.90 GiBKV cache 6.00 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 (31.48 GiB): Apple M4 Pro, A100 40GB, NVIDIA L40S, RTX 6000 Ada
FP8 (19.17 GiB): NVIDIA L4, Apple M4, NVIDIA A30, RTX 3090 / Ti
INT4 (13.40 GiB): NVIDIA T4, RTX 4060 Ti 16GB, RTX 4070 Ti Super, RTX 4080 / Super

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	2	6
RTX 5090	does not fit	7	11
RTX 6000 Ada	9	17	21
H100 SXM 80GB	28	36	40
RTX PRO 6000 Blackwell	37	46	50
AMD Ryzen AI Max+ 395 (128GB)	38	46	50
H200 141GB	65	73	77
B200 180GB	88	96	100

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

Fine-tuning Qwen2.5 14B Instruct

Method	Memory	Smallest GPU
Full fine-tune	222.0 GiB	AMD MI325X
LoRA (rank 16, attention)	29.78 GiB	RTX 5090
QLoRA (rank 16, attention)	9.36 GiB	RTX 3060

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 14B Instruct that is 192.0 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.