How much VRAM does Mistral-Medium-3.5-128B need?

Q4_K_M is exactly 74,897,139,136 bytes (69.75 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is Mistral-Medium-3.5-128B's KV cache?

11.00 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

Mistral-Medium-3.5-128B

mistralai/Mistral-Medium-3.5-128B

Mistral-Medium-3.5-128B at Q4_K_M is exactly 74,897,139,136 bytes (69.75 GiB / 74.90 GB) — an effective 4.692 bits per weight, not the nominal 4. Its KV cache at 32K is 11.00 GiB.

From the file· summed from 3 file(s)From the file· KV per layer

Parameters

128B

Architecture

mistral3

88 layers

Context

262,144

native (config.json)

License

other

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
I1-IQ1_S	25.33 GiB	27,193,745,664	1.704	—	mradermacher
I1-IQ1_M	27.59 GiB	29,620,281,600	1.856	—	mradermacher
I1-IQ2_XXS	31.35 GiB	33,664,508,160	2.109	—	mradermacher
UD-IQ2_XXS	32.54 GiB	34,937,742,432	2.189	—	unsloth
IQ1_S	32.92 GiB	35,344,326,496	2.214	—	bartowski
IQ1_M	34.69 GiB	37,251,817,312	2.334	—	bartowski
I1-IQ2_XS	34.75 GiB	37,315,125,504	2.338	—	mradermacher
I1-IQ2_S	37.01 GiB	39,740,875,008	2.490	—	mradermacher
IQ2_XXS	37.65 GiB	40,430,968,672	2.533	—	bartowski
I1-IQ2_M	40.02 GiB	42,976,256,256	2.692	—	mradermacher
I1-Q2_K_S	40.05 GiB	43,000,635,648	2.694	—	mradermacher
IQ2_XS	40.41 GiB	43,389,525,856	2.718	—	bartowski
UD-IQ2_M	41.08 GiB	44,109,112,416	2.763	—	unsloth
IQ2_S	42.90 GiB	46,065,753,952	2.886	—	bartowski
I1-Q2_K	43.39 GiB	46,590,697,728	2.919	—	mradermacher
Q2_K	43.39 GiB	46,590,698,592	2.919	—	unsloth
Q2_K_L	43.74 GiB	46,968,185,952	2.942	—	unsloth
I1-IQ3_XXS	45.04 GiB	48,365,674,752	3.030	—	mradermacher
IQ2_M	45.27 GiB	48,609,075,040	3.045	—	bartowski
UD-IQ3_XXS	45.87 GiB	49,252,770,912	3.085	—	unsloth
Q2_K	46.44 GiB	49,860,681,568	3.123	—	bartowski
Q2_K_L2 shards	47.90 GiB	51,433,545,760	3.222	—	bartowski
I1-IQ3_XS	48.11 GiB	51,659,251,968	3.236	—	mradermacher
IQ3_XXS2 shards	48.59 GiB	52,172,005,408	3.268	—	bartowski
I1-Q3_K_S	50.63 GiB	54,366,937,344	3.406	—	mradermacher
Q3_K_S3 shards	50.63 GiB	54,366,938,528	3.406	—	unsloth
I1-IQ3_S	50.77 GiB	54,514,000,128	3.415	—	mradermacher
IQ3_XS2 shards	51.16 GiB	54,929,236,000	3.441	—	bartowski
I1-IQ3_M	52.89 GiB	56,793,473,280	3.558	—	mradermacher
Q3_K_S2 shards	53.05 GiB	56,962,162,720	3.568	—	bartowski
IQ3_M2 shards	55.44 GiB	59,527,110,688	3.729	—	bartowski
I1-Q3_K_M	56.46 GiB	60,619,858,176	3.797	—	mradermacher
Q3_K_M3 shards	56.46 GiB	60,619,859,360	3.797	—	unsloth
Q3_K_M2 shards	58.94 GiB	63,281,143,840	3.964	—	bartowski
I1-Q3_K_L	61.53 GiB	66,071,404,800	4.139	—	mradermacher
Q3_K_L2 shards	62.40 GiB	67,005,685,792	4.198	—	bartowski
I1-IQ4_XS	62.47 GiB	67,074,105,600	4.202	—	mradermacher
IQ4_XS3 shards	62.51 GiB	67,124,438,432	4.205	—	unsloth
IQ4_XS2 shards	64.39 GiB	69,141,635,104	4.331	—	bartowski
IQ4_NL3 shards	66.03 GiB	70,896,166,304	4.441	—	unsloth

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	1.38 GiB	1.38 GiB	—	88 / 0 / 0
8,192	2.75 GiB	2.75 GiB	—	88 / 0 / 0
16,384	5.50 GiB	5.50 GiB	—	88 / 0 / 0
32,768	11.00 GiB	11.00 GiB	—	88 / 0 / 0
65,536	22.00 GiB	22.00 GiB	—	88 / 0 / 0
131,072	44.00 GiB	44.00 GiB	—	88 / 0 / 0

Compare with

same modality, comparable size

vs Qwen3-Coder-30B-A3B-Instruct vs Qwen3.6-27B vs Qwen3.8-27B vs gemma-4-E4B-it vs Qwen3-30B-A3B-Thinking-2507 vs Qwen3-8B 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 66.90 GiB. The real file is 69.75 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

12288

Vocab

131,072

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does Mistral-Medium-3.5-128B need?: Q4_K_M is exactly 74,897,139,136 bytes (69.75 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is Mistral-Medium-3.5-128B's KV cache?: 11.00 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 Mistral-Medium-3.5-128B 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.