LLMs · IBM

Granite 4.0 H Small (32B-A9B)

Name: Granite 4.0 H Small (32B-A9B)
Author: IBM

Parameters: 32B (9.0B active)
Architecture: Mixture of experts
Max context: 128K tokens
License: apache-2.0
Released: 2025-09

How to run it

Quickest · Ollama

One command downloads a ready-to-run version and starts a chat.

ollama run granite4:small-h

Get Ollama at ollama.com/download (opens in a new tab) · model page (opens in a new tab)

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

huggingface.co/ibm-granite/granite-4.0-h-small-GGUF (opens in a new tab)

ollama run hf.co/ibm-granite/granite-4.0-h-small-GGUF

Original weights · Hugging Face

Full-precision files from IBM, for fine-tuning, vLLM, Transformers.

huggingface.co/ibm-granite/granite-4.0-h-small (opens in a new tab)

No sign-in needed.

Official repo · GitHub

Reference code, inference scripts and docs from IBM. The weights themselves download from Hugging Face.

github.com/ibm-granite/granite-4.0-language-models (opens in a new tab)

git clone https://github.com/ibm-granite/granite-4.0-language-models

IBM’s official Granite 4.0 repo.

Links checked 5 days ago against Hugging Face, Ollama and GitHub.

Memory needed

Weights plus KV cache and runtime overhead. Leave a little extra free for the operating system and other apps.

KV cache: hybrid Mamba-Transformer: only the attention layers keep a KV cache.

Quantization	Quality	Weights	Total at 8K context	Total at 32K context	Download
`Q2_K`	Low	12.9 GB	13.6 GB	14.0 GB	GGUF (opens in a new tab)
`Q3_K_M`	Low	15.8 GB	16.5 GB	16.9 GB	GGUF (opens in a new tab)
`Q4_K_M`	Good	19.6 GB	20.4 GB	20.8 GB	GGUF (opens in a new tab)
`Q5_K_M`	High	22.9 GB	23.6 GB	24.0 GB	GGUF (opens in a new tab)
`Q6_K`	High	26.4 GB	27.1 GB	27.5 GB	GGUF (opens in a new tab)
`Q8_0`	Very high	34.1 GB	34.9 GB	35.3 GB	GGUF (opens in a new tab)
`F16`	Full	64.4 GB	65.2 GB	65.6 GB	GGUF (opens in a new tab) Original weights (opens in a new tab)

Best hardware for Granite 4.0 H Small (32B-A9B)

The least expensive GPU in our catalog that runs it without offloading is the AMD Radeon RX 7600 XT (launch price $329) at Q2_K.

Grade · points	Hardware	Quant	Memory	Speed	Fit
82 ptsGreat (A) · 82 of 100 points: Model quality 27.1/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 15.0/15	NVIDIA GeForce RTX 5090	`Q5_K_M`	25.2 GB / 32.0 GB	101 tok/s	Runs well
80 ptsGreat (A) · 80 of 100 points: Model quality 26.5/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15	AMD Radeon RX 7900 XTX	`Q4_K_M`	21.6 GB / 24.0 GB	56.4 tok/s	Tight fit
80 ptsGreat (A) · 80 of 100 points: Model quality 26.5/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15	NVIDIA GeForce RTX 3090	`Q4_K_M`	21.6 GB / 24.0 GB	63.8 tok/s	Tight fit
80 ptsGreat (A) · 80 of 100 points: Model quality 26.5/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15	NVIDIA GeForce RTX 4090	`Q4_K_M`	21.6 GB / 24.0 GB	68.7 tok/s	Tight fit
77 ptsGreat (A) · 77 of 100 points: Model quality 27.9/45, Memory fit 25.0/25, Speed 15.0/15, Memory use 9.5/15	NVIDIA RTX PRO 6000 Blackwell	`Q8_0`	39.7 GB / 96.0 GB	62.2 tok/s	Runs well
76 ptsGreat (A) · 76 of 100 points: Model quality 26.5/45, Memory fit 25.0/25, Speed 13.4/15, Memory use 11.6/15	Apple M4 Max (64 GB)	`Q4_K_M`	24.2 GB / 48.0 GB	32.1 tok/s	Runs well
75 ptsGreat (A) · 75 of 100 points: Model quality 26.5/45, Memory fit 25.0/25, Speed 9.5/15, Memory use 14.9/15	Apple M4 Pro (48 GB)	`Q4_K_M`	23.3 GB / 36.0 GB	16.0 tok/s	Runs well
74 ptsGreat (A) · 74 of 100 points: Model quality 26.5/45, Memory fit 25.0/25, Speed 11.5/15, Memory use 11.6/15	Apple M1 Max (64 GB)	`Q4_K_M`	24.2 GB / 48.0 GB	23.5 tok/s	Runs well
73 ptsGreat (A) · 73 of 100 points: Model quality 19.5/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15	AMD Radeon RX 9070 XT	`Q2_K`	14.4 GB / 16.0 GB	61.7 tok/s	Tight fit
73 ptsGreat (A) · 73 of 100 points: Model quality 19.5/45, Memory fit 24.0/25, Speed 15.0/15, Memory use 15.0/15	NVIDIA GeForce RTX 4070 Ti SUPER	`Q2_K`	14.4 GB / 16.0 GB	75.3 tok/s	Tight fit

Check it on your own hardware

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.