Ministral-3-3B-Reasoning-2512 · unsloth

Ministral-3-3B-Reasoning-2512 IQ4_XS

This file is exactly 1,959,276,288 bytes — 1.82 GiB / 1.96 GB at an effective 3.687 bits per weight. The nominal rate for IQ4_XS is lower; mixed-precision tensors make the real figure higher, always.

From the file· 1 file(s), summedFrom the file· 236 tensors parsed

Get it

1.82 GiB · 1 file
llama.cpp
llama-cli -hf unsloth/Ministral-3-3B-Reasoning-2512-GGUF:IQ4_XS

Downloads and runs in one step, resolving the quantization by name.

Hugging Face CLI
hf download unsloth/Ministral-3-3B-Reasoning-2512-GGUF Ministral-3-3B-Reasoning-2512-IQ4_XS.gguf
Direct download

Straight from the Hugging Face CDN — we host nothing and earn nothing from this. Verify what you received against the exact byte count above; a size mismatch is the usual cause of a file that will not load.

Repository →
Size
1.82 GiB
1.96 GB
Effective bpw
3.687
from real bytes ÷ params
Tensors
236
Header
7.88 MB
GGUF metadata

What this quantization actually contains

per-tensor types, parsed from the GGUF header
IQ4_XS
156
F32
53
Q5_K
26
Q6_K
1

A quantization label names a mixture, not a uniform precision. Attention and output tensors are routinely kept at higher precision than the label implies, which is exactly why the effective bits-per-weight above exceeds the nominal rate.

Fit by accelerator

AcceleratorMemory○Total @4K◐Total @32K◐Fits◐tok/s @4K◐
Arc A310 4GB4 GB3.04 GiB5.89 GiByes4K only29±30%
Arc A350M 4GB4 GB3.04 GiB5.89 GiByes4K only27±30%
Arc A370M 4GB4 GB3.04 GiB5.89 GiByes4K only27±30%
Arc A530M 4GB4 GB3.04 GiB5.89 GiByes4K only50±30%
Arc Pro A30M 4GB4 GB3.04 GiB5.89 GiByes4K only27±30%
Radeon Pro W64004 GB3.14 GiB5.99 GiByes4K only35±26.5%
Radeon RX 64004 GB3.14 GiB5.99 GiByes4K only35±26.5%
Radeon RX 6500 XT4 GB3.14 GiB5.99 GiByes4K only40±26.5%
RTX A4004 GB3.24 GiB6.09 GiByes4K only24±22%
Arc A380 6GB6 GB3.04 GiB5.89 GiByes4K only43±30%
Arc Pro A40 6GB6 GB3.04 GiB5.89 GiByes4K only44±30%
Arc Pro A50 6GB6 GB3.04 GiB5.89 GiByes4K only44±30%
GeForce RTX 20606 GB3.04 GiB5.89 GiByes4K only102±12.9%
GeForce RTX 30506 GB3.04 GiB5.89 GiByes4K only53±12.9%
GeForce RTX 3060 OEM6 GB3.04 GiB5.89 GiByes4K only102±12.9%
RTX A20006 GB3.24 GiB6.09 GiByes4K only71±22%
Apple M18 GB2.79 GiB5.64 GiByes23±8.3%
Apple M28 GB2.79 GiB5.64 GiByes32±8.3%
Apple M38 GB2.79 GiB5.64 GiByes32±8.3%
Apple M48 GB2.79 GiB5.64 GiByes36±8.3%
Arc A530M 8GB8 GB3.04 GiB5.89 GiByes50±30%
Arc A550M 8GB8 GB3.04 GiB5.89 GiByes50±30%
Arc A570M 8GB8 GB3.04 GiB5.89 GiByes50±30%
Arc A580 8GB8 GB3.04 GiB5.89 GiByes102±30%
○Spec sheet◐Predictedwhat these mean

Memory at context

ContextWeights●KV (f16)●KV (q8_0)●Working buffer◐Total (f16)◐
4,0961.82 GiB0.41 GiB0.22 GiB0.31 GiB2.54 GiB
8,1921.82 GiB0.81 GiB0.43 GiB0.31 GiB2.95 GiB
16,3841.82 GiB1.63 GiB0.86 GiB0.31 GiB3.76 GiB
32,7681.82 GiB3.25 GiB1.73 GiB0.31 GiB5.39 GiB
65,5361.82 GiB6.50 GiB3.45 GiB0.31 GiB8.64 GiB
131,0721.82 GiB13.00 GiB6.91 GiB0.31 GiB15.14 GiB
●From the file◐Predictedwhat these mean

Quantizing the KV cache to q8_0 is a roughly 2× lever on the dominant term at long context, and it is the single most useful setting most local users never touch. Totals here exclude the allocator reserve your driver takes, which is hardware-specific — the per-accelerator table above includes it.