OLMo 2 13B Instruct memory requirements

13.7B parameters — 25.5 GiB of weights at FP16, 7.8 GiB at Q4_K_M, and 800.0 KiB of KV cache per token on top. Solved against 22 real accelerators; 21 of them can hold it.

Open the calculator →All models

Parameters

13.7B

13,716,198,400 exactly

Weights, Q4_K_M

7.8 GiB

25.5 GiB unquantized

KV cache

800.0 KiB

per token · 40L × 40 KV × 128

Smallest card (Q4)

12 GB

GeForce RTX 3060 12GB

Weights at each precision

Precision	Bits/weight	Weights	Source	Notes
FP16 / BF16	16.00	25.5 GiB	measured file	Unquantized. Two bytes per parameter, exactly.
Q8_0	8.50	13.6 GiB	measured file	Effectively lossless; the usual reference point for quantized quality.
Q6_K	6.57	10.5 GiB	measured file	Quality loss is hard to measure on most benchmarks.
Q5_K_M	5.69	9.1 GiB	measured file	A middle point when Q4 fits with too little room for context.
Q4_K_M	4.87	7.8 GiB	measured file	The default choice for local inference — the best size/quality knee.
Q3_K_M	3.95	6.3 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	4,096	27 tok/s
H200 141GB (SXM)	141 GB	4800 GB/s	FP16 / BF16	4,096	156 tok/s
Apple M4 Max (128GB)	128 GB	546 GB/s	FP16 / BF16	4,096	18 tok/s
H100 80GB (SXM5)	80 GB	3350 GB/s	FP16 / BF16	4,096	109 tok/s
A100 80GB (SXM)	80 GB	2039 GB/s	FP16 / BF16	4,096	66 tok/s
Jetson AGX Orin (64GB)	64 GB	204.8 GB/s	FP16 / BF16	4,096	7 tok/s
RTX 6000 Ada Generation	48 GB	960 GB/s	FP16 / BF16	4,096	31 tok/s
L40S	48 GB	864 GB/s	FP16 / BF16	4,096	28 tok/s
Apple M4 Pro (48GB)	48 GB	273 GB/s	FP16 / BF16	4,096	9 tok/s
A100 40GB (SXM)	40 GB	1555 GB/s	FP16 / BF16	4,096	51 tok/s
GeForce RTX 5090	32 GB	1792 GB/s	FP16 / BF16	4,096	58 tok/s
GeForce RTX 4090	24 GB	1008 GB/s	Q8_0	4,096	56 tok/s
Radeon RX 7900 XTX	24 GB	960 GB/s	Q8_0	4,096	54 tok/s
GeForce RTX 3090	24 GB	936 GB/s	Q8_0	4,096	52 tok/s
Apple M4 (24GB)	24 GB	120 GB/s	Q8_0	4,096	7 tok/s
GeForce RTX 5080	16 GB	960 GB/s	Q6_K	4,096	66 tok/s
GeForce RTX 5070 Ti	16 GB	896 GB/s	Q6_K	4,096	61 tok/s
GeForce RTX 4080 SUPER	16 GB	736 GB/s	Q6_K	4,096	50 tok/s
GeForce RTX 4070 Ti SUPER	16 GB	672 GB/s	Q6_K	4,096	46 tok/s
GeForce RTX 4060 Ti 16GB	16 GB	288 GB/s	Q6_K	4,096	20 tok/s
GeForce RTX 3060 12GB	12 GB	360 GB/s	Q4_K_M	4,096	31 tok/s
Jetson Orin Nano Super (8GB)	8 GB	102 GB/s	no fit	—	—

Questions about this model

How much VRAM does OLMo 2 13B Instruct need?

25.5 GiB for the weights at FP16 and 7.8 GiB at Q4_K_M, from an exact count of 13,716,198,400 parameters. On top of that, the KV cache costs 800.0 KiB per token of context — 6.3 GiB at 8k and 25.0 GiB at 32k.

What is the smallest GPU that runs OLMo 2 13B Instruct?

Of the 22 accelerators here, the smallest that holds Q4_K_M weights is the GeForce RTX 3060 12GB at 12 GB — 7.8 GiB of weights against 11.0 GiB usable, leaving room for 4k tokens. For unquantized weights you need at least a GeForce RTX 5090.

Why is the KV cache for OLMo 2 13B Instruct the size it is?

Because it stores one key and one value vector per token, per layer: 40 layers × 40 KV heads × 128 dimensions × 2 (K and V) × 2 bytes = 800.0 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.