How much VRAM does SmolLM2-135M-Instruct need?

Q4_K_M is exactly 105,454,144 bytes (0.10 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is SmolLM2-135M-Instruct's KV cache?

0.70 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

HuggingFaceTB · text

SmolLM2-135M-Instruct

HuggingFaceTB/SmolLM2-135M-Instruct

SmolLM2-135M-Instruct at Q4_K_M is exactly 105,454,144 bytes (0.10 GiB / 0.11 GB) — an effective 6.272 bits per weight, not the nominal 4. Its KV cache at 32K is 0.70 GiB.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

135M

Architecture

llama

30 layers

Context

8,192

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q2_K	0.08 GiB	88,201,792	5.246	—	unsloth
Q2_K	0.08 GiB	88,202,080	5.246	—	bartowski
Q3_K_S	0.08 GiB	88,202,080	5.246	—	bartowski
Q2_K_L	0.08 GiB	88,202,080	5.246	—	bartowski
IQ3_XS	0.08 GiB	88,202,080	5.246	—	bartowski
Q2_K	0.08 GiB	88,202,208	5.246	—	mradermacher
Q3_K_S	0.08 GiB	88,202,208	5.246	—	mradermacher
IQ3_M	0.08 GiB	90,213,472	5.365	—	bartowski
IQ4_XS	0.08 GiB	90,897,760	5.406	272	bartowski
IQ4_XS	0.09 GiB	91,312,608	5.431	—	mradermacher
Q4_0	0.09 GiB	91,893,088	5.465	272	bartowski
Q3_K_M	0.09 GiB	93,510,208	5.561	272	unsloth
Q3_K_M	0.09 GiB	93,510,496	5.561	272	bartowski
Q3_K_M	0.09 GiB	93,510,624	5.561	—	mradermacher
Q3_K_L	0.09 GiB	97,532,992	5.801	—	lmstudio-community
Q3_K_L	0.09 GiB	97,533,280	5.801	—	bartowski
Q3_K_L	0.09 GiB	97,533,408	5.801	—	mradermacher
Q4_K_S	0.10 GiB	102,039,904	6.069	—	bartowski
Q4_K_S	0.10 GiB	102,040,032	6.069	—	mradermacher
Q4_K_M	0.10 GiB	105,454,144	6.272	272	lmstudio-community
Q4_K_M	0.10 GiB	105,454,144	6.272	—	unsloth
Q4_K_L	0.10 GiB	105,454,432	6.272	—	bartowski
Q4_K_M	0.10 GiB	105,454,432	6.272	272	bartowski
Q4_K_M	0.10 GiB	105,454,560	6.272	—	mradermacher
Q5_K_S	0.10 GiB	109,974,880	6.540	—	bartowski
Q5_K_S	0.10 GiB	109,975,008	6.540	—	mradermacher
Q5_K_M	0.10 GiB	112,103,488	6.667	272	unsloth
Q5_K_M	0.10 GiB	112,103,776	6.667	272	bartowski
Q5_K_L	0.10 GiB	112,103,776	6.667	—	bartowski
Q5_K_M	0.10 GiB	112,103,904	6.667	—	mradermacher
Q6_K	0.13 GiB	138,382,912	8.230	—	unsloth
Q6_K	0.13 GiB	138,382,912	8.230	272	lmstudio-community
Q6_K	0.13 GiB	138,383,200	8.230	272	bartowski
Q6_K_L	0.13 GiB	138,383,200	8.230	—	bartowski
Q6_K	0.13 GiB	138,383,328	8.230	—	mradermacher
Q8_0	0.13 GiB	144,811,072	8.612	272	unsloth
Q8_0	0.13 GiB	144,811,072	8.612	272	lmstudio-community
Q8_0	0.13 GiB	144,811,360	8.612	272	bartowski
Q8_0	0.13 GiB	144,811,488	8.612	—	mradermacher
F16	0.25 GiB	270,885,952	16.110	272	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.09 GiB	0.09 GiB	—	30 / 0 / 0
8,192	0.18 GiB	0.18 GiB	—	30 / 0 / 0
16,384	0.35 GiB	0.35 GiB	—	30 / 0 / 0
32,768	0.70 GiB	0.70 GiB	—	30 / 0 / 0
65,536	1.41 GiB	1.41 GiB	—	30 / 0 / 0
131,072	2.81 GiB	2.81 GiB	—	30 / 0 / 0

Compare with

same modality, comparable size

vs ced-base vs Qwen3-0.6B vs gemma-3-270m-it vs tinygemma3_cifar vs LFM2.5-230M vs LFM2.5-350M All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 0.07 GiB. The real file is 0.10 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

Hidden size

576

Vocab

49,152

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does SmolLM2-135M-Instruct need?: Q4_K_M is exactly 105,454,144 bytes (0.10 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is SmolLM2-135M-Instruct's KV cache?: 0.70 GiB at 32K context with an f16 cache, computed per layer. Quantizing the cache to q8_0 roughly halves it, which is often the difference between a context length fitting and not.
Which quantization of SmolLM2-135M-Instruct should I use?: Q4_K_M is the usual default. Pick the largest quantization that fits your card at the context you actually need — the table above gives exact sizes for every one published.