mistralai · text

Ministral-8B-Instruct-2410

mistralai/Ministral-8B-Instruct-2410

Ministral-8B-Instruct-2410 at Q4_K_M is exactly 4,911,500,096 bytes (4.57 GiB / 4.91 GB) — an effective 4.899 bits per weight, not the nominal 4. Its KV cache at 32K is 4.50 GiB, not the 4.50 GiB a flat formula predicts.

From the file· summed from 1 file(s)From the file· KV per layer
Parameters
8.0B
Architecture
llama
36 layers
Context
32,768
native (config.json)
License
other

Shipped quantizations

exact bytes, summed from published files
QuantSizeExact bytesEffective bpwTensorsPublisher
IQ2_M2.76 GiB2,958,330,6882.951bartowski
Q2_K2.97 GiB3,185,478,4643.178bartowski
IQ3_XS3.28 GiB3,521,940,2883.513bartowski
Q3_K_S3.41 GiB3,664,677,6963.656bartowski
Q2_K_L3.45 GiB3,709,766,4643.701bartowski
IQ3_M3.53 GiB3,791,686,4643.782bartowski
Q3_K_M3.74 GiB4,019,227,4564.009bartowski
Q3_K_L4.03 GiB4,326,460,2244.316bartowski
IQ4_XS4.14 GiB4,448,226,1124.437bartowski
Q4_04.34 GiB4,658,464,9924.647vanta-research
Q4_04.35 GiB4,671,048,5124.660bartowski
Q4_K_S4.36 GiB4,685,728,5764.674bartowski
Q4_K_M4.57 GiB4,911,500,0964.899bartowski
Q4_K_L4.95 GiB5,309,958,9765.297bartowski
Q5_K_S5.21 GiB5,593,795,3925.580bartowski
Q5_K_M5.33 GiB5,724,146,4965.710bartowski
Q5_K_L5.64 GiB6,055,496,5126.040bartowski
Q6_K6.14 GiB6,587,583,2966.571bartowski
Q6_K_L6.38 GiB6,847,630,1446.831bartowski
Q8_07.94 GiB8,529,808,1928.509bartowski
F1614.95 GiB16,048,097,82416.009bartowski

KV cache by context

computed per layer — this model uses sliding-window attention
ContextKV cache (f16)Flat formulaOverstated byFull / windowed / recurrent
4,0960.56 GiB0.56 GiB9 / 27 / 0
8,1921.13 GiB1.13 GiB9 / 27 / 0
16,3842.25 GiB2.25 GiB9 / 27 / 0
32,7684.50 GiB4.50 GiB9 / 27 / 0
65,5365.68 GiB9.00 GiB1.59×9 / 27 / 0
131,0727.93 GiB18.00 GiB2.27×9 / 27 / 0

27 of 36 layers cache only a 32,768-token window rather than the full context, on a period of . Figures assume the default configuration; --swa-full disables the saving entirely.

Compare with

same modality, comparable size

Will it run on your card?

full quant x context sweep

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 4.20 GiB. The real file is 4.57 GiB, because a quantization is a mixture and some tensors are always kept at higher precision. The larger discrepancy is the cache: a flat formula gives 4.50 GiB at 32K context where the real figure is 4.50 GiB, because most of this model's layers cache a fixed window rather than the whole context.

Architecture

from config.json
Layers
36
Attention heads
32
KV heads
8
Head dim
128
Hidden size
4096
Vocab
131,072
Sliding window
32768
SWA period
MLA
no
Experts
Experts per token
use_sliding_window

Questions people ask

How much VRAM does Ministral-8B-Instruct-2410 need?
Q4_K_M is exactly 4,911,500,096 bytes (4.57 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-8B-Instruct-2410's KV cache?
4.50 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-8B-Instruct-2410 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.