granite-speech-4.1-2b · handy-computer

granite-speech-4.1-2b Q8_0

This file is exactly 2,559,878,848 bytes — 2.38 GiB / 2.56 GB at an effective 8.853 bits per weight. The nominal rate for Q8_0 is lower; mixed-precision tensors make the real figure higher, always.

From the file· 1 file(s), summedFrom the file· 940 tensors parsed

Get it

2.38 GiB · 1 file
llama.cpp
llama-cli -hf handy-computer/granite-speech-4.1-2b-gguf:Q8_0

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

Hugging Face CLI
hf download handy-computer/granite-speech-4.1-2b-gguf granite-speech-4.1-2b-Q8_0.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
2.38 GiB
2.56 GB
Effective bpw
8.853
from real bytes ÷ params
Tensors
940
Header
3.61 MB
GGUF metadata

What this quantization actually contains

per-tensor types, parsed from the GGUF header
F32
473
Q8_0
434
F16
33

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 GB3.48 GiB5.67 GiByes4K only24±30%
Arc A350M 4GB4 GB3.48 GiB5.67 GiByes4K only22±30%
Arc A370M 4GB4 GB3.48 GiB5.67 GiByes4K only22±30%
Arc A530M 4GB4 GB3.48 GiB5.67 GiByes4K only42±30%
Arc Pro A30M 4GB4 GB3.48 GiB5.67 GiByes4K only22±30%
Radeon Pro W64004 GB3.58 GiB5.77 GiByes4K only30±26.5%
Radeon RX 64004 GB3.58 GiB5.77 GiByes4K only30±26.5%
Radeon RX 6500 XT4 GB3.58 GiB5.77 GiByes4K only33±26.5%
RTX A4004 GB3.68 GiB5.87 GiByes4K only20±22%
Arc A380 6GB6 GB3.48 GiB5.67 GiByes4K only36±30%
Arc Pro A40 6GB6 GB3.48 GiB5.67 GiByes4K only37±30%
Arc Pro A50 6GB6 GB3.48 GiB5.67 GiByes4K only37±30%
GeForce RTX 20606 GB3.48 GiB5.67 GiByes4K only86±12.9%
GeForce RTX 30506 GB3.48 GiB5.67 GiByes4K only44±12.9%
GeForce RTX 3060 OEM6 GB3.48 GiB5.67 GiByes4K only86±12.9%
RTX A20006 GB3.68 GiB5.87 GiByes4K only59±22%
Apple M18 GB3.23 GiB5.42 GiByes19±8.3%
Apple M28 GB3.23 GiB5.42 GiByes27±8.3%
Apple M38 GB3.23 GiB5.42 GiByes27±8.3%
Apple M48 GB3.23 GiB5.42 GiByes31±8.3%
Arc A530M 8GB8 GB3.48 GiB5.67 GiByes42±30%
Arc A550M 8GB8 GB3.48 GiB5.67 GiByes42±30%
Arc A570M 8GB8 GB3.48 GiB5.67 GiByes42±30%
Arc A580 8GB8 GB3.48 GiB5.67 GiByes88±30%
○Spec sheet◐Predictedwhat these mean

Memory at context

ContextWeights●KV (f16)●KV (q8_0)●Working buffer◐Total (f16)◐
4,0962.38 GiB0.31 GiB0.17 GiB0.28 GiB2.98 GiB
8,1922.38 GiB0.63 GiB0.33 GiB0.28 GiB3.29 GiB
16,3842.38 GiB1.25 GiB0.66 GiB0.28 GiB3.92 GiB
32,7682.38 GiB2.50 GiB1.33 GiB0.28 GiB5.17 GiB
65,5362.38 GiB5.00 GiB2.66 GiB0.28 GiB7.67 GiB
131,0722.38 GiB10.00 GiB5.31 GiB0.28 GiB12.67 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.