Skip to content

Granite 4.0 H Small VRAM requirements

How much VRAM Granite 4.0 H Small needs
4-bit weights: 18.2 GBWith 8K context: 19.1 GBWith 32K context: 19.5 GB8-bit, 8K context: 32.8 GB
Parameters
32.2B
Active per token
8.8B
Longest context
128K
Released
Sep 2025
License
Apache 2.0
Strongest at
Not rated

How much VRAM Granite 4.0 H Small needs

Memory Granite 4.0 H Small needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit60.0 GB60.9 GB61.3 GB62.8 GB
8-bit31.9 GB32.8 GB33.1 GB34.6 GB
6-bit24.6 GB25.5 GB25.9 GB27.4 GB
5-bit21.3 GB22.2 GB22.6 GB24.1 GB
4-bit18.2 GB19.1 GB19.5 GB21.0 GB
3-bit14.7 GB15.6 GB15.9 GB17.4 GB
2-bit11.2 GB12.1 GB12.5 GB14.0 GB
1.6-bit6.2 GB7.1 GB7.5 GB9.0 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Granite 4.0 H Small

The best way Granite 4.0 H Small runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Granite 4.0 H Small on RTX 4060 report
RTX 5060Too bigPDF: Granite 4.0 H Small on RTX 5060 report
RTX 3070Too bigPDF: Granite 4.0 H Small on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Granite 4.0 H Small on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: Granite 4.0 H Small on RTX 3060 12 GB report
Arc B580Too bigPDF: Granite 4.0 H Small on Arc B580 report
RTX 4070Too bigPDF: Granite 4.0 H Small on RTX 4070 report
RTX 4070 SuperToo bigPDF: Granite 4.0 H Small on RTX 4070 Super report
RTX 5070Too bigPDF: Granite 4.0 H Small on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: Granite 4.0 H Small on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: Granite 4.0 H Small on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: Granite 4.0 H Small on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: Granite 4.0 H Small on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: Granite 4.0 H Small on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: Granite 4.0 H Small on RTX 4080 Super report
RTX 5070 TiToo bigPDF: Granite 4.0 H Small on RTX 5070 Ti report
RTX 5080Too bigPDF: Granite 4.0 H Small on RTX 5080 report
Radeon RX 7900 XTX5-bit44K69 tok/s2.8 sPDF: Granite 4.0 H Small on Radeon RX 7900 XTX report
RTX 30905-bit44K84 tok/s1.8 sPDF: Granite 4.0 H Small on RTX 3090 report
RTX 40905-bit44K89 tok/s0.8 sPDF: Granite 4.0 H Small on RTX 4090 report
Radeon AI PRO R97006-bit128K45 tok/s1.8 sPDF: Granite 4.0 H Small on Radeon AI PRO R9700 report
RTX 50906-bit128K123 tok/s0.6 sPDF: Granite 4.0 H Small on RTX 5090 report
2× RTX 30908-bit128K60 tok/s1.0 sPDF: Granite 4.0 H Small on 2× RTX 3090 report
2× RTX 40908-bit128K64 tok/s0.4 sPDF: Granite 4.0 H Small on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit128K81 tok/s0.9 sPDF: Granite 4.0 H Small on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit128K61 tok/s0.7 sPDF: Granite 4.0 H Small on RTX 6000 Ada report
L40S8-bit128K56 tok/s1.1 sPDF: Granite 4.0 H Small on L40S report
2× RTX 50908-bit128K101 tok/s0.3 sPDF: Granite 4.0 H Small on 2× RTX 5090 report
4× RTX 30908-bit128K60 tok/s0.5 sPDF: Granite 4.0 H Small on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit128K101 tok/s0.5 sPDF: Granite 4.0 H Small on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: Granite 4.0 H Small on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: Granite 4.0 H Small on Mac mini M4 24 GB report
Mac mini M4 32 GB4-bit96K17 tok/s40 sPDF: Granite 4.0 H Small on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: Granite 4.0 H Small on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit112K21 tok/s20 sPDF: Granite 4.0 H Small on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit128K21 tok/s20 sPDF: Granite 4.0 H Small on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit128K35 tok/s10.0 sPDF: Granite 4.0 H Small on M4 Max 128 GB report
M5 Max 128 GB8-bit128K39 tok/s2.9 sPDF: Granite 4.0 H Small on M5 Max 128 GB report
M3 Ultra 96 GB8-bit128K43 tok/s5.6 sPDF: Granite 4.0 H Small on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K43 tok/s5.6 sPDF: Granite 4.0 H Small on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K43 tok/s5.6 sPDF: Granite 4.0 H Small on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K57 tok/s1.7 sPDF: Granite 4.0 H Small on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit128K16 tok/s5.7 sPDF: Granite 4.0 H Small on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit128K16 tok/s5.7 sPDF: Granite 4.0 H Small on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K21 tok/s2.0 sPDF: Granite 4.0 H Small on RTX Spark report
DGX Spark8-bit128K19 tok/s2.0 sPDF: Granite 4.0 H Small on DGX Spark report
Dual DGX Spark8-bit128K19 tok/s1.1 sPDF: Granite 4.0 H Small on Dual DGX Spark report
DGX Station GB3008-bit128K234 tok/s0.2 sPDF: Granite 4.0 H Small on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K101 tok/s0.3 sPDF: Granite 4.0 H Small on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K101 tok/s0.1 sPDF: Granite 4.0 H Small on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K101 tok/s0.1 sPDF: Granite 4.0 H Small on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit128K112 tok/s1.3 sPDF: Granite 4.0 H Small on A100 80 GB SXM report
H100 SXM8-bit128K157 tok/s0.4 sPDF: Granite 4.0 H Small on H100 SXM report
4× H200 NVL8-bit128K193 tok/s0.1 sPDF: Granite 4.0 H Small on 4× H200 NVL report
8× H200 NVL8-bit128K193 tok/s0.1 sPDF: Granite 4.0 H Small on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf ibm-granite/granite-4.0-h-small-GGUF
vLLM
vllm serve ibm-granite/granite-4.0-h-small
MLX (Mac)
mlx_lm.chat --model mlx-community/granite-4.0-h-small-4bit

Questions

How much VRAM does Granite 4.0 H Small need?

Granite 4.0 H Small needs 18.2 GB for its weights at 4-bit, 19.1 GB with an 8K context and 19.5 GB with 32K. At 8-bit, 32.8 GB.

What is the minimum VRAM for Granite 4.0 H Small?

About 15.6 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the Radeon RX 7900 XTX.

How big is Granite 4.0 H Small?

32.2B parameters (8.8B active per token): 60.0 GB of weights at 16-bit, 18.2 GB at 4-bit.

Can Granite 4.0 H Small run on an RTX 4090?

Yes. At 5-bit it holds 44K of context and writes about 89 tokens a second.

More models