How much VRAM does Ministral-3-3B-Instruct-2512 need?

Q4_K_M is exactly 2,146,497,184 bytes (2.00 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Ministral-3-3B-Instruct-2512's KV cache?

3.25 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.

mistralai · text

Ministral-3-3B-Instruct-2512

mistralai/Ministral-3-3B-Instruct-2512

Ministral-3-3B-Instruct-2512 at Q4_K_M is exactly 2,146,497,184 bytes (2.00 GiB / 2.15 GB) — an effective 4.461 bits per weight, not the nominal 4. Its KV cache at 32K is 3.25 GiB.

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

Parameters

3.8B

Architecture

mistral3

26 layers

Context

262,144

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
UD-IQ1_S	0.90 GiB	969,159,968	2.014	—	unsloth
UD-IQ1_M	0.95 GiB	1,019,196,704	2.118	—	unsloth
UD-IQ2_XXS	1.03 GiB	1,108,948,256	2.305	—	unsloth
UD-IQ2_M	1.25 GiB	1,339,888,928	2.785	—	unsloth
UD-IQ3_XXS	1.36 GiB	1,457,804,576	3.030	—	unsloth
Q2_K_L	1.36 GiB	1,458,959,648	3.032	—	unsloth
Q2_K	1.36 GiB	1,458,959,648	3.032	—	unsloth
Q3_K_S	1.53 GiB	1,639,150,880	3.407	—	unsloth
Q3_K_M	1.67 GiB	1,795,552,544	3.732	—	unsloth
IQ4_XS	1.82 GiB	1,959,277,856	4.072	—	unsloth
Q4_0	1.91 GiB	2,046,375,200	4.253	—	unsloth
IQ4_NL	1.91 GiB	2,051,290,400	4.263	—	unsloth
Q4_K_S	1.91 GiB	2,053,256,480	4.268	—	unsloth
Q4_K_M	2.00 GiB	2,146,497,184	4.461	—	AmarettoLabs
Q4_K_M	2.00 GiB	2,146,497,824	4.461	—	unsloth
Q4_K_M	2.00 GiB	2,147,023,008	4.462	—	mistralai
Q4_1	2.08 GiB	2,230,203,680	4.635	—	unsloth
Q5_K_S	2.25 GiB	2,419,340,576	5.028	—	unsloth
Q5_K_M	2.30 GiB	2,473,652,896	5.141	—	AmarettoLabs
Q5_K_M	2.30 GiB	2,473,653,536	5.141	—	unsloth
Q5_K_M	2.30 GiB	2,474,178,720	5.142	—	mistralai
Q6_K	2.63 GiB	2,821,255,840	5.864	—	AmarettoLabs
Q6_K	2.63 GiB	2,821,256,480	5.864	—	unsloth
Q8_0	3.40 GiB	3,651,678,880	7.590	—	AmarettoLabs
Q8_0	3.40 GiB	3,651,679,520	7.590	—	unsloth
Q8_0	3.40 GiB	3,652,204,704	7.591	—	mistralai
BF16	6.39 GiB	6,866,220,032	14.271	—	unsloth
BF16	6.40 GiB	6,866,745,504	14.272	—	mistralai

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.41 GiB	0.41 GiB	—	26 / 0 / 0
8,192	0.81 GiB	0.81 GiB	—	26 / 0 / 0
16,384	1.63 GiB	1.63 GiB	—	26 / 0 / 0
32,768	3.25 GiB	3.25 GiB	—	26 / 0 / 0
65,536	6.50 GiB	6.50 GiB	—	26 / 0 / 0
131,072	13.00 GiB	13.00 GiB	—	26 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-4B vs Qwen3-8B vs HyperCLOVAX-SEED-Text-Instruct-1.5B vs Llama-3.2-1B-Instruct vs Qwen3-VL-8B-Instruct-abliterated-v1 vs Llama-3.1-8B-Instruct 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 2.02 GiB. The real file is 2.00 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

128

Hidden size

3072

Vocab

131,072

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Ministral-3-3B-Instruct-2512 need?: Q4_K_M is exactly 2,146,497,184 bytes (2.00 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Ministral-3-3B-Instruct-2512's KV cache?: 3.25 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 Ministral-3-3B-Instruct-2512 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.