Llama-3.2-3B-Instruct · unsloth

Llama-3.2-3B-Instruct UD-IQ3_XXS

This file is exactly 1,370,147,264 bytes — 1.28 GiB / 1.37 GB at an effective 3.412 bits per weight. The nominal rate for UD-IQ3_XXS is lower; mixed-precision tensors make the real figure higher, always.

From the file· 1 file(s), summed

Get it

1.28 GiB · 1 file
llama.cpp
llama-cli -hf unsloth/Llama-3.2-3B-Instruct-GGUF:UD-IQ3_XXS

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

Hugging Face CLI
hf download unsloth/Llama-3.2-3B-Instruct-GGUF Llama-3.2-3B-Instruct-UD-IQ3_XXS.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
1.28 GiB
1.37 GB
Effective bpw
3.412
from real bytes ÷ params
Tensors
—
Header
—
GGUF metadata

Fit by accelerator

AcceleratorMemory○Total @4K◐Total @32K◐Fits◐tok/s @4K◐
Arc A310 4GB4 GB2.52 GiB5.58 GiByes4K only37±30%
Arc A350M 4GB4 GB2.52 GiB5.58 GiByes4K only34±30%
Arc A370M 4GB4 GB2.52 GiB5.58 GiByes4K only34±30%
Arc A530M 4GB4 GB2.52 GiB5.58 GiByes4K only64±30%
Arc Pro A30M 4GB4 GB2.52 GiB5.58 GiByes4K only34±30%
Radeon Pro W64004 GB2.62 GiB5.68 GiByes4K only45±26.5%
Radeon RX 64004 GB2.62 GiB5.68 GiByes4K only45±26.5%
Radeon RX 6500 XT4 GB2.62 GiB5.68 GiByes4K only51±26.5%
RTX A4004 GB2.72 GiB5.78 GiByes4K only32±22%
Arc A380 6GB6 GB2.52 GiB5.58 GiByes4K only54±30%
Arc Pro A40 6GB6 GB2.52 GiB5.58 GiByes4K only56±30%
Arc Pro A50 6GB6 GB2.52 GiB5.58 GiByes4K only56±30%
GeForce RTX 20606 GB2.52 GiB5.58 GiByes4K only130±12.9%
GeForce RTX 30506 GB2.52 GiB5.58 GiByes4K only68±12.9%
GeForce RTX 3060 OEM6 GB2.52 GiB5.58 GiByes4K only130±12.9%
RTX A20006 GB2.72 GiB5.78 GiByes4K only92±22%
Apple M18 GB2.27 GiB5.33 GiByes28±8.3%
Apple M28 GB2.27 GiB5.33 GiByes39±8.3%
Apple M38 GB2.27 GiB5.33 GiByes39±8.3%
Apple M48 GB2.27 GiB5.33 GiByes44±8.3%
Arc A530M 8GB8 GB2.52 GiB5.58 GiByes64±30%
Arc A550M 8GB8 GB2.52 GiB5.58 GiByes64±30%
Arc A570M 8GB8 GB2.52 GiB5.58 GiByes64±30%
Arc A580 8GB8 GB2.52 GiB5.58 GiByes125±30%
○Spec sheet◐Predictedwhat these mean

Memory at context

ContextWeights●KV (f16)●KV (q8_0)●Working buffer◐Total (f16)◐
4,0961.28 GiB0.44 GiB0.23 GiB0.31 GiB2.02 GiB
8,1921.28 GiB0.88 GiB0.46 GiB0.31 GiB2.46 GiB
16,3841.28 GiB1.75 GiB0.93 GiB0.31 GiB3.33 GiB
32,7681.28 GiB3.50 GiB1.86 GiB0.31 GiB5.08 GiB
65,5361.28 GiB7.00 GiB3.72 GiB0.31 GiB8.58 GiB
131,0721.28 GiB14.00 GiB7.44 GiB0.31 GiB15.58 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.