[whatmodel.sh](../../README.md) / Granite 4.0 H Small (32B-A9B)
# Granite 4.0 H Small (32B-A9B)
IBM · released 2025-09 · 32.2B params (9B active) · apache-2.0
**score 52.5** ±0.6 · #327 of 732 · from 39 benchmark results across 7 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks Granite 4.0 H Small (32B-A9B) among every model on that task or test, worst to best.
task vs the others rank counts tests
agent tasks ▄·▄········● #1 of 4* 0.7% 1
coding █···▄······● #1 of 4 15% 7
chat & writing ▁▁▁▃▆▄▆●▆▇▇▃ #238 of 402 6.6% 3
math ▁▁▃▅█●▄█▄▂▁▁ #234 of 431 22% 8
agent tasks · #1 of 4 · 0.7% of its score
test result vs the others in its run difficulty counts why
BFCL v3 (maker) 64.7% ▁▁▁▁▁▃▅█▆●▄▁ beats 123/171 medium · avg 60% 0.7% blunt · at its level
coding · #1 of 4 · 15% of its score
test result vs the others in its run difficulty counts why
HumanEval (maker) 88.0% ▁▁▁▁▂▂▆▆█▅●▁ beats 195/205 easy · avg 84% 3.8% sharp
HumanEval+ (maker) 83.0% ▁·▁▁▂▃▄▄█●▄▃ beats 114/145 easy · avg 79% 3.6% sharp
MBPP (maker) 84.0% ▁▁▁▄▅█▆▅▄▁●▁ beats 82/86 easy · avg 78% 1.8% typical
BigCodeBench (maker) 46.2% ▅····▃·█▃▅█● beats 12/12 medium · avg 41% 1.7% at its level
MultiPL-E (maker) 57.4% ▃··▄·▅●▇▄█▃▄ beats 12/31 medium · avg 56% 1.6% at its level
MBPP+ (maker) 71.0% ▁··▁▁▃▄█●▄▁▁ beats 87/120 easy · avg 68% 1.3% at its level
CRUXEval-O (maker) 50.2% ▂▂██▅●▁·▁·▁▁ beats 45/51 medium · avg 46% 0.9% blunt · at its level
chat & writing · #238 of 402 · 6.6% of its score
Beats [Mistral Small 4 (119B-A6.5B)](mistral-small-4.md), [Qwen2.5 Coder 32B Instruct](qwen2-5-coder-32b.md), [Ministral 3 14B Reasoning 2512](ministral-3-14b-reasoning.md) here, though they rank higher overall. Trails [Flan-T5 XXL](flan-t5-xxl.md), [LFM2.5 1.2B Instruct](lfm2-5-1-2b-instruct.md), [Gemma 3 4B](gemma-3-4b.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
LMArena (text) 1287.3 ▁▁▁▂▃▃●▆▆▇█▃ beats 117/385 avg 1260 4.3% blunt
IFEval prompt-level strict (maker) 85.2% ▁▁··▂··▂▇●▃▃ beats 14/26 easy · avg 80% 1.1% typical
IFEval instruction-level strict (… 89.9% ▄▄····▄▄█▄·● beats 7/8 very easy · avg 88% 1.1% typical
math · #234 of 431 · 22% of its score
Beats [GPT-4o (2024-08-06)](gpt-4o-2024-08-06.md), [Command A (03-2025)](command-a.md), [GPT-4o (2024-05-13)](gpt-4o-2024-05-13.md) here, though they rank higher overall. Trails [Gemma 7B (base)](gemma-7b-base.md), [Phi-2 (2.7B)](phi-2.md), [Granite 8B Code Base (4K)](granite-8b-code-base-4k.md) here, though they rank lower overall.
test result vs the others in its run difficulty counts why
GSM8K 83.9% █····●··▄·▄█ beats 2/7 very easy · avg 86% 7.6% sharp · several runs
DeepMind Mathematics (maker) 59.3% ▅····▅·█●▃·▅ beats 7/10 medium · avg 53% 2.7% at its level
Minerva MATH (maker) 74.0% ▁·▃▁▁▂▁█▄▁▃● beats 25/27 easy · avg 68% 2.6% at its level
MATH (subset unspecified) 75.7% █●··▄█▄····▄ beats 2/7 easy · avg 80% 2.4% typical
GSM-Symbolic (maker) 87.4% ▃▃····▃·▅▅█● beats 10/10 easy · avg 82% 1.9% typical
PolyMath (Marco evaluation) 28.9% ▄··▄█●··▄··▄ beats 4/7 hard · avg 29% 1.8% at its level
MGSM (Marco evaluation) 75.7% ▄·▄·▄▄▄·●··█ beats 5/7 easy · avg 69% 1.5% at its level
MGSM, 5 languages (en, es, fr, ja… 38.7% █·▄▄●█···█·▄ beats 5/10 medium · avg 44% 1.0% at its level
other tests (17)
test result vs the others in its run difficulty counts why
MMLU 76.7% ▄···▄▄·▄●··█ beats 5/7 easy · avg 74% 8.3% at its level · several runs
Medmarks-V mean win rate (19 veri… 0.4452 ▁▄▃▃●▄█▃▅▁▁▁ beats 10/30 avg 0.418 8.0% sharp
GPQA Diamond 38.6% ▄···█·▄·●·▄▄ beats 4/7 hard · avg 37% 7.4% sharp
MMLU-Pro 57.1% ▄····█·▄●·▄▄ beats 4/7 medium · avg 54% 6.5% at its level · several runs
GPQA (subset unspecified) (maker) 40.6% ▂▁▂▂▄▃●▃▅█▃▃ beats 40/106 very hard · avg 35% 4.8% sharp
Medmarks-V weighted mean win rate… 44.5% ▂▃▅▂●▃█▅█▅▃▆ beats 10/34 medium · avg 41% 3.5% blunt · at its level
BIG-Bench Hard (maker) 81.6% ▁▁▁▁▂▄▅▆█●▃▁ beats 137/155 easy · avg 74% 3.4% sharp · at its level
MMLU-ProX-Lite (Marco) 51.6% █··▄··█··●·▄ beats 6/7 medium · avg 42% 2.8% at its level
INCLUDE (maker) 64.0% ▂····▂▂▄●▄█▄ beats 14/36 medium · avg 57% 2.3% at its level
GlobalMMLU (Marco evaluation) 67.4% ▃··▃▃··▅···● beats 5/7 medium · avg 56% 2.0% at its level
MMMLU (maker) 69.7% ▂▁▁▂▃▂▄●▅██▃ beats 19/46 medium · avg 62% 2.0% at its level
MMMLU (Marco evaluation) 68.1% ▄···▄·█·▄·●█ beats 5/7 medium · avg 59% 1.7% at its level
AGIEval (Marco suite unspecified) 44.7% ▄··▄●▄··▄··█ beats 2/7 medium · avg 46% 1.4% blunt · at its level
DiagnosticIQ industrial maintenan… 53.2% ▁··▃▁●▁▅█▅▁▅ beats 4/24 hard · avg 51% 0.9% blunt · at its level
EuroEval HellaSwag (English), tes… 84.8% ▃·▁▁▂▁▂▅█●▄▁ beats 83/100 medium · avg 67% 0.8% sharp · at its level · noisy
EuroEval Life in the UK (English … 86.5% ▁▁▁▁▁▁▁▂▄●▄▃ beats 103/146 easy · avg 82% 0.5% noisy
Vectara hallucination leaderboard… 94.8% ▁▂▁▂▅█▂▃▄●▁▁ beats 88/98 very easy · avg 90% 0.0% blunt
source results share of the score
ibm-granite/granite-4.0-h-small model c… 21 48% ▇▇▇▇▇▇
Marco-MoE: Open Multilingual Mixture-of… 11 34% ▇▇▇▇
Medmarks: A Comprehensive Open-Source L… 2 12% ▇
LMArena leaderboard dataset 1 4% ▇
EuroEval leaderboards 2 1% ▇
DiagnosticIQ: A Benchmark for LLM-Based… 1 1% ▇
Vectara hallucination leaderboard 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 52.5
Licences: ibm-granite/granite-4.0-h-small model card ([Apache-2.0](../../licenses/ibm-granite--granite-4.0-h-small.txt)) · Marco-MoE: Open Multilingual Mixture-of-Expert Language Models with Efficient Upcycling ([CC BY 4.0](../../licenses/cards-final-NOTICE.txt)) · Medmarks: A Comprehensive Open-Source LLM Benchmark Suite for Medical Tasks ([CC BY 4.0](../../licenses/paper-2605.01417v1-CC-BY-4.0.txt)) · LMArena leaderboard dataset ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · EuroEval leaderboards ([MIT](https://github.com/EuroEval/leaderboards/blob/1a77f1d2f8ba1ba05ab73287fb090d76bf41b2e0/LICENSE)) · DiagnosticIQ: A Benchmark for LLM-Based Industrial Maintenance Action Recommendation from Symbolic Rules ([CC BY 4.0](https://creativecommons.org/licenses/by/4.0/)) · Vectara hallucination leaderboard ([Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0))
## How to run it
### Through an API
api (Replicate) · $0.11/M blended · $0.06 in / $0.25 out per 1M tokens · [replicate.com](https://replicate.com/ibm-granite/granite-4.0-h-small)
[price source](https://replicate.com/ibm-granite/granite-4.0-h-small): USD per 1M tokens, Replicate official model ibm-granite/granite-4.0-h-small, billed per input/output token (billing table on the model page; A100 80GB backing hardware not billed separately for this official model). Currency: USD. Replicate official model page, observed 2026-09-25. · [context source](https://huggingface.co/ibm-granite/granite-4.0-h-small/blob/b8c0982bab7fde4eb48110f5a069527c008fab39/config.json): Configured sequence length, corroborated by the official card architecture table. Dense Micro and hybrid H-Small are distinct models.
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run Granite 4.0 H Small (32B-A9B) (Q4_K_M, needs ~19.5 GB, from [huggingface.co/ibm-granite](https://huggingface.co/ibm-granite/granite-4.0-h-small-GGUF/tree/6522095069a86fc3186b8632850e4373dd850cac))
```sh
llama cli -hf ibm-granite/granite-4.0-h-small-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf ibm-granite/granite-4.0-h-small-GGUF:Q4_K_M
```
Other versions: Q8_0 (~34.3 GB) `llama cli -hf ibm-granite/granite-4.0-h-small-GGUF:Q8_0`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run Granite 4.0 H Small (32B-A9B) (Q4_K_M, needs ~19.5 GB, from [huggingface.co/ibm-granite](https://huggingface.co/ibm-granite/granite-4.0-h-small-GGUF/tree/6522095069a86fc3186b8632850e4373dd850cac))
```sh
lms get https://huggingface.co/ibm-granite/granite-4.0-h-small-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~34.3 GB) `lms get https://huggingface.co/ibm-granite/granite-4.0-h-small-GGUF@Q8_0 -y` · 4bit (~18.1 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/granite-4.0-h-small-MLX-4bit -y` · 8bit (~34.2 GB, Mac only) `lms get https://huggingface.co/lmstudio-community/granite-4.0-h-small-MLX-8bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run Granite 4.0 H Small (32B-A9B) (Q4_K_M, needs ~19.82 GB, from [ollama.com](https://ollama.com/library/granite4/tags))
```sh
ollama run granite4:small-h
```
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run Granite 4.0 H Small (32B-A9B) (4bit, needs ~18.1 GB, Mac only, from [huggingface.co/lmstudio-community](https://huggingface.co/lmstudio-community/granite-4.0-h-small-MLX-4bit/tree/6418bb430b1166392e7cd881d58928f97672b5f4))
```sh
mlx_lm.chat --model lmstudio-community/granite-4.0-h-small-MLX-4bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model lmstudio-community/granite-4.0-h-small-MLX-4bit
```
Other versions: 8bit (~34.2 GB, Mac only) `mlx_lm.chat --model lmstudio-community/granite-4.0-h-small-MLX-8bit`
Model repository and documentation: [huggingface.co/ibm-granite](https://huggingface.co/ibm-granite/granite-4.0-h-small)
## Quantization evidence
How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent. They don't describe every download of that precision.
precision test full → quantized kept
FP8 (IBM official ibm-granite… GSM8K 5-shot strict-match (lm-eva… 0.852 → 0.853 100.1%
FP8 (IBM official ibm-granite… MMLU 5-shot acc (lm-evaluation-ha… 0.771 → 0.772 100.0%
FP8 (IBM official ibm-granite… OpenLLM v1 average of 6 tasks (AR… 0.766 → 0.766 100.0%
FP8 (IBM official ibm-granite… IFEval 0-shot instruction-level s… 0.875 → 0.872 99.6%
FP8 (IBM official ibm-granite… MMLU-Pro 5-shot (OpenLLM v2 confi… 0.48 → 0.476 99.2%
FP8 (IBM official ibm-granite… OpenLLM v2 average of 6 tasks (IF… 0.542 → 0.536 98.8%
FP8 block-wise (llm-compresso… GSM8K 5-shot strict-match (lm-eva… 0.852 → 0.855 100.4%
FP8 block-wise (llm-compresso… MMLU 5-shot acc (lm-evaluation-ha… 0.771 → 0.772 100.1%
FP8 block-wise (llm-compresso… OpenLLM v1 average of 6 tasks (AR… 0.766 → 0.766 100.0%
FP8 block-wise (llm-compresso… MMLU-Pro 5-shot (OpenLLM v2 confi… 0.48 → 0.479 99.9%
FP8 block-wise (llm-compresso… IFEval 0-shot instruction-level s… 0.875 → 0.867 99.0%
FP8 block-wise (llm-compresso… OpenLLM v2 average of 6 tasks (IF… 0.542 → 0.533 98.4%
FP8 dynamic (llm-compressor F… MMLU 5-shot acc (lm-evaluation-ha… 0.771 → 0.772 100.0%
FP8 dynamic (llm-compressor F… IFEval 0-shot instruction-level s… 0.875 → 0.874 99.9%
FP8 dynamic (llm-compressor F… OpenLLM v1 average of 6 tasks (AR… 0.766 → 0.765 99.9%
FP8 dynamic (llm-compressor F… GSM8K 5-shot strict-match (lm-eva… 0.852 → 0.848 99.6%
FP8 dynamic (llm-compressor F… MMLU-Pro 5-shot (OpenLLM v2 confi… 0.48 → 0.476 99.1%
FP8 dynamic (llm-compressor F… OpenLLM v2 average of 6 tasks (IF… 0.542 → 0.53 97.8%
Each kept value links to its source.
[interactive version](https://whatmodel.sh/model/granite-4-0-h-small) · [JSON](https://whatmodel.sh/api?format=json&all=1)