How much VRAM does granite-speech-4.1-2b need?

Q4_K_M is exactly 1,139,247,200 bytes (1.06 GiB) in weights. Add the KV cache, which depends on context length, plus roughly half a gigabyte of runtime overhead.

How large is granite-speech-4.1-2b's KV cache?

2.50 GiB at 32K context with an f16 cache, computed per layer from the model's architecture. Quantizing the cache to q8_0 roughly halves it.

ibm-granite · audio asr

granite-speech-4.1-2b

ibm-granite/granite-speech-4.1-2b

granite-speech-4.1-2b at Q4_K_M is exactly 1,139,247,200 bytes (1.06 GiB / 1.14 GB) — an effective 3.940 bits per weight, not the nominal 4. Its KV cache at 32K is 2.50 GiB.

From the file· summed from 1 file(s)From the file· KV per layer

Parameters

2.3B

Architecture

granite_speech

40 layers

Context

4,096

native (config.json)

License

apache-2.0

Shipped quantizations

● exact bytes, summed from published files

Quant	Size●	Exact bytes●	Effective bpw●	Tensors●	Publisher
Q4_K_M	1.06 GiB	1,139,247,200	3.940	—	ibm-granite
Q5_K_M	1.23 GiB	1,319,602,528	4.564	—	ibm-granite
Q6_K	1.41 GiB	1,511,230,080	5.226	—	ibm-granite
Q4_K_M	1.49 GiB	1,602,904,800	5.543	940	handy-computer
Q5_K_M	1.70 GiB	1,829,704,544	6.328	940	handy-computer
Q8_0	1.82 GiB	1,956,154,944	6.765	—	ibm-granite
Q6_K	1.89 GiB	2,024,967,936	7.003	940	handy-computer
Q8_0	2.38 GiB	2,559,878,848	8.853	940	handy-computer
Q4_K	2.74 GiB	2,941,043,168	10.171	—	cstr
BF16	3.43 GiB	3,678,444,864	12.722	—	ibm-granite
BF16	4.31 GiB	4,631,640,064	16.018	—	handy-computer
F16	4.31 GiB	4,632,623,104	16.021	940	handy-computer
F16	5.20 GiB	5,581,887,616	19.304	—	cstr

KV cache by context

computed per layer

Context	KV cache (f16)●	Flat formula	Overstated by	Full / windowed / recurrent
4,096	0.31 GiB	0.31 GiB	—	40 / 0 / 0
8,192	0.63 GiB	0.63 GiB	—	40 / 0 / 0
16,384	1.25 GiB	1.25 GiB	—	40 / 0 / 0
32,768	2.50 GiB	2.50 GiB	—	40 / 0 / 0
65,536	5.00 GiB	5.00 GiB	—	40 / 0 / 0
131,072	10.00 GiB	10.00 GiB	—	40 / 0 / 0

Measured

published by a third party, attributed below

Metric	Value	What it means
rtf	546.8
RTFx	546.8	Higher is better — audio seconds processed per second of compute.
Word error rate	8.23%	Lower is better — the share of words transcribed incorrectly.
Word error rate	7.06%	Lower is better — the share of words transcribed incorrectly.
Word error rate	8.18%	Lower is better — the share of words transcribed incorrectly.
Word error rate	3.47%	Lower is better — the share of words transcribed incorrectly.
Word error rate	7.72%	Lower is better — the share of words transcribed incorrectly.
Word error rate	4.18%	Lower is better — the share of words transcribed incorrectly.
Word error rate	2.17%	Lower is better — the share of words transcribed incorrectly.
Word error rate	5.40%	Lower is better — the share of words transcribed incorrectly.
Word error rate	1.02%	Lower is better — the share of words transcribed incorrectly.
Word error rate	4.90%	Lower is better — the share of words transcribed incorrectly.

Benchmarked· by open-asr-leaderboard-english-short-latest

RTFx measured by the Open ASR Leaderboard on a single datacenter GPU at a large batch size. It ranks models against each other; it says nothing about throughput on consumer hardware. We reproduce these figures with attribution; they are not ours and we have not verified the runs. Source: open-asr-leaderboard-english-short-latest.

Compare with

same modality, comparable size

vs nemotron-3.5-asr-streaming-0.6b vs parakeet-unified-en-0.6b vs cohere-transcribe-03-2026 vs parakeet-tdt-0.6b-v3 vs whisper-medium vs Voxtral-Mini-4B-Realtime-2602 All alternatives →Will it fit on my card? →

Why other calculators give a different number

A parameters × bits ÷ 8 estimate puts Q4_K_M at roughly 1.21 GiB. The real file is 1.06 GiB, because a quantization is a mixture and some tensors are always kept at higher precision.

Architecture

from config.json

Layers

Attention heads

KV heads

Head dim

128

Hidden size

2048

Vocab

100,353

Sliding window

none

SWA period

—

MLA

Experts

—

Experts per token

—

use_sliding_window

—

Questions people ask

How much VRAM does granite-speech-4.1-2b need?: Q4_K_M is exactly 1,139,247,200 bytes (1.06 GiB) in weights. Add the KV cache, which depends on your context length, plus roughly half a gigabyte of runtime overhead.
How large is granite-speech-4.1-2b's KV cache?: 2.50 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 granite-speech-4.1-2b 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.