Phi-3.5-mini 3.8B Instruct memory requirements

3.8B parameters — 7.1 GiB of weights at FP16, 2.2 GiB at Q4_K_M, and 384.0 KiB of KV cache per token on top. Solved against 22 real accelerators; 22 of them can hold it.

Open the calculator →All models

Parameters

3.8B

3,821,079,552 exactly

Weights, Q4_K_M

2.2 GiB

7.1 GiB unquantized

KV cache

384.0 KiB

per token · 32L × 32 KV × 96

Smallest card (Q4)

8 GB

Jetson Orin Nano Super (8GB)

Weights at each precision

Precision	Bits/weight	Weights	Source	Notes
FP16 / BF16	16.00	7.1 GiB	measured file	Unquantized. Two bytes per parameter, exactly.
Q8_0	8.50	3.8 GiB	measured file	Effectively lossless; the usual reference point for quantized quality.
Q6_K	6.57	2.9 GiB	measured file	Quality loss is hard to measure on most benchmarks.
Q5_K_M	5.89	2.6 GiB	measured file	A middle point when Q4 fits with too little room for context.
Q4_K_M	5.01	2.2 GiB	measured file	The default choice for local inference — the best size/quality knee.
Q3_K_M	4.09	1.8 GiB	measured file	Measurable quality loss. Worth it only to make a model fit at all.

Every accelerator

Largest quant that fits with at least 4k of context, the context it leaves, and the bandwidth ceiling on decode speed. Each row links to the worked page for that pairing.

Accelerator	Memory	Bandwidth	Best quant	Max context	Ceiling
Apple M3 Ultra (512GB)	512 GB	819 GB/s	FP16 / BF16	131,072	75 tok/s
H200 141GB (SXM)	141 GB	4800 GB/s	FP16 / BF16	131,072	442 tok/s
Apple M4 Max (128GB)	128 GB	546 GB/s	FP16 / BF16	131,072	50 tok/s
H100 80GB (SXM5)	80 GB	3350 GB/s	FP16 / BF16	131,072	308 tok/s
A100 80GB (SXM)	80 GB	2039 GB/s	FP16 / BF16	131,072	188 tok/s
Jetson AGX Orin (64GB)	64 GB	204.8 GB/s	FP16 / BF16	111,636	19 tok/s
RTX 6000 Ada Generation	48 GB	960 GB/s	FP16 / BF16	101,151	88 tok/s
L40S	48 GB	864 GB/s	FP16 / BF16	101,151	80 tok/s
Apple M4 Pro (48GB)	48 GB	273 GB/s	FP16 / BF16	78,868	25 tok/s
A100 40GB (SXM)	40 GB	1555 GB/s	FP16 / BF16	81,053	143 tok/s
GeForce RTX 5090	32 GB	1792 GB/s	FP16 / BF16	60,955	165 tok/s
GeForce RTX 4090	24 GB	1008 GB/s	FP16 / BF16	40,858	93 tok/s
Radeon RX 7900 XTX	24 GB	960 GB/s	FP16 / BF16	40,858	88 tok/s
GeForce RTX 3090	24 GB	936 GB/s	FP16 / BF16	40,858	86 tok/s
Apple M4 (24GB)	24 GB	120 GB/s	FP16 / BF16	29,716	11 tok/s
GeForce RTX 5080	16 GB	960 GB/s	FP16 / BF16	20,760	88 tok/s
GeForce RTX 5070 Ti	16 GB	896 GB/s	FP16 / BF16	20,760	82 tok/s
GeForce RTX 4080 SUPER	16 GB	736 GB/s	FP16 / BF16	20,760	68 tok/s
GeForce RTX 4070 Ti SUPER	16 GB	672 GB/s	FP16 / BF16	20,760	62 tok/s
GeForce RTX 4060 Ti 16GB	16 GB	288 GB/s	FP16 / BF16	20,760	27 tok/s
GeForce RTX 3060 12GB	12 GB	360 GB/s	FP16 / BF16	10,711	33 tok/s
Jetson Orin Nano Super (8GB)	8 GB	102 GB/s	Q8_0	6,055	16 tok/s

Questions about this model

How much VRAM does Phi-3.5-mini 3.8B Instruct need?

7.1 GiB for the weights at FP16 and 2.2 GiB at Q4_K_M, from an exact count of 3,821,079,552 parameters. On top of that, the KV cache costs 384.0 KiB per token of context — 3.0 GiB at 8k and 12.0 GiB at 32k.

What is the smallest GPU that runs Phi-3.5-mini 3.8B Instruct?

Of the 22 accelerators here, the smallest that holds Q4_K_M weights is the Jetson Orin Nano Super (8GB) at 8 GB — 2.2 GiB of weights against 6.0 GiB usable, leaving room for 10k tokens. For unquantized weights you need at least a GeForce RTX 3060 12GB.

Why is the KV cache for Phi-3.5-mini 3.8B Instruct the size it is?

Because it stores one key and one value vector per token, per layer: 32 layers × 32 KV heads × 96 dimensions × 2 (K and V) × 2 bytes = 384.0 KiB per token. This model uses full multi-head attention, so every attention head keeps its own K and V — the cache is as large as the architecture allows. None of this can be read off the parameter count.

Compare with

Method and limits. The parameter count and every architecture figure on this page are read from the model's own published config and the Hub's index over its tensor shapes, fetched 2026-08-07 — nothing here is recalled from a model's name. Weight bytes are the size of a real published file wherever one exists and the parameter count times the published llama.cpp bits-per-weight where it does not; every row says which. Speed figures are roofline bounds — memory bandwidth divided by bytes read per token — not benchmarks. Nothing on this page is written by a language model.