Qwen3-4B-Instruct-2507 · unsloth

Qwen3-4B-Instruct-2507 Q5_K_M

This file is exactly 2,889,514,080 bytes — 2.69 GiB / 2.89 GB at an effective 5.747 bits per weight. The nominal rate for Q5_K_M is lower; mixed-precision tensors make the real figure higher, always.

From the file· 1 file(s), summedFrom the file· 398 tensors parsed

Get it

2.69 GiB · 1 file
llama.cpp
llama-cli -hf unsloth/Qwen3-4B-Instruct-2507-GGUF:Q5_K_M

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

Hugging Face CLI
hf download unsloth/Qwen3-4B-Instruct-2507-GGUF Qwen3-4B-Instruct-2507-Q5_K_M.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
2.69 GiB
2.89 GB
Effective bpw
5.747
from real bytes ÷ params
Tensors
398
Header
5.96 MB
GGUF metadata

What this quantization actually contains

per-tensor types, parsed from the GGUF header
Q5_K
216
F32
145
Q6_K
37

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 GB4.06 GiB8.00 GiBno—
Arc A350M 4GB4 GB4.06 GiB8.00 GiBno—
Arc A370M 4GB4 GB4.06 GiB8.00 GiBno—
Arc A530M 4GB4 GB4.06 GiB8.00 GiBno—
Arc Pro A30M 4GB4 GB4.06 GiB8.00 GiBno—
Radeon Pro W64004 GB4.16 GiB8.10 GiBno—
Radeon RX 64004 GB4.16 GiB8.10 GiBno—
Radeon RX 6500 XT4 GB4.16 GiB8.10 GiBno—
RTX A4004 GB4.26 GiB8.20 GiBno—
Arc A380 6GB6 GB4.06 GiB8.00 GiByes4K only30±30%
Arc Pro A40 6GB6 GB4.06 GiB8.00 GiByes4K only31±30%
Arc Pro A50 6GB6 GB4.06 GiB8.00 GiByes4K only31±30%
GeForce RTX 20606 GB4.06 GiB8.00 GiByes4K only72±12.9%
GeForce RTX 30506 GB4.06 GiB8.00 GiByes4K only37±12.9%
GeForce RTX 3060 OEM6 GB4.06 GiB8.00 GiByes4K only72±12.9%
RTX A20006 GB4.26 GiB8.20 GiByes4K only50±22%
Apple M18 GB3.81 GiB7.75 GiByes4K only16±8.3%
Apple M28 GB3.81 GiB7.75 GiByes4K only23±8.3%
Apple M38 GB3.81 GiB7.75 GiByes4K only23±8.3%
Apple M48 GB3.81 GiB7.75 GiByes4K only27±8.3%
Arc A530M 8GB8 GB4.06 GiB8.00 GiByes4K only36±30%
Arc A550M 8GB8 GB4.06 GiB8.00 GiByes4K only36±30%
Arc A570M 8GB8 GB4.06 GiB8.00 GiByes4K only36±30%
Arc A580 8GB8 GB4.06 GiB8.00 GiByes4K only75±30%
○Spec sheet◐Predictedwhat these mean

Memory at context

ContextWeights●KV (f16)●KV (q8_0)●Working buffer◐Total (f16)◐
4,0962.69 GiB0.56 GiB0.30 GiB0.31 GiB3.56 GiB
8,1922.69 GiB1.13 GiB0.60 GiB0.31 GiB4.13 GiB
16,3842.69 GiB2.25 GiB1.20 GiB0.31 GiB5.25 GiB
32,7682.69 GiB4.50 GiB2.39 GiB0.31 GiB7.50 GiB
65,5362.69 GiB9.00 GiB4.78 GiB0.31 GiB12.00 GiB
131,0722.69 GiB18.00 GiB9.56 GiB0.31 GiB21.00 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.