BestLLMfor EN Your hardware. Your LLM. Your call.
APIOpen data Find my LLM
Model fiche

Granite 4.0 H-Small 32B-A9B

By IBM · United States

Updated 2026-07-13

chat general moe
Parameters
32B
License
Apache 2.0
Context
125k
VRAM (Q4)
19 GB
Released
October 2025

Overview

IBM's hybrid Mamba-2 + MoE model with 32B total and 9B active parameters, engineered to slash long-context memory use by roughly 70% versus comparable transformers under Apache 2.0.

When to pick this model

  • Long-document RAG pipelines where VRAM is the bottleneck
  • Enterprise deployments needing a permissive Apache 2.0 license
  • Self-hosted assistants handling 100k+ token transcripts
  • Cost-sensitive inference at sustained high concurrency
  • Workloads where you want MoE throughput without the H100-class footprint

VRAM requirements by quantization

VRAM REQUIRED (GB)81216243248Q4_K_M19 GBQ5_K_M23 GBQ8_035 GBFP1664 GB
QuantizationVRAM required
Q4_K_M (recommended)19 GB
Q5_K_M23 GB
Q8_035 GB
FP16 (no quantization)64 GB

VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.

In practice, Granite 4.0 H-Small 32B-A9B wants a 24 GB card at Q4_K_M (19 GB). Stepping up to Q8_0 nearly doubles the footprint to 35 GB, and unquantized FP16 weights take 64 GB — plan your GPU around the Q4 or Q5 figure unless you specifically need the higher fidelity.

Without a GPU, Granite 4.0 H-Small 32B-A9B needs roughly 32 GB of system RAM to run on CPU via llama.cpp or Ollama — workable for background jobs, but far slower than GPU inference. Throughput estimates from our compatibility engine: around 10 tokens/sec on entry-level GPUs, on the order of 30 tokens/sec on a mid-range card, and up to 75 tokens/sec on high-end hardware — assuming the chosen quantization fully fits in VRAM.

What hardware do you need

The table below matches Granite 4.0 H-Small 32B-A9B to common GPU memory tiers, using the highest-fidelity quantization that fully fits each card class. Spilling layers to system RAM works but costs most of the speed, so size your card to the quantization you actually want to run.

GPU memoryExample cardsBest fit for Granite 4.0 H-Small 32B-A9B
8 GBRTX 5070 Laptop, RTX 5060, RTX 5060 Ti 8GBDoes not fit — needs 19 GB at Q4_K_M
12 GBRTX 5070, RTX 5070 Ti Laptop, RTX 4080 LaptopDoes not fit — needs 19 GB at Q4_K_M
16 GBRTX 5080, RTX 4080 Super, Radeon RX 9070 XTDoes not fit — needs 19 GB at Q4_K_M
24 GBRTX 4090, Radeon RX 7900 XTX, RTX 5090 LaptopQ5_K_M (23 GB used)
32 GBRTX 5090Q5_K_M (23 GB used)

Which GPU should you buy to run Granite 4.0 H-Small 32B-A9B?

To run Granite 4.0 H-Small 32B-A9B locally at Q4, you need ~19 GB of VRAM. The best value for this is a RTX 4090 (24 GB VRAM).

Check RTX 4090 price on Amazon →

As an Amazon Associate, BestLLMfor earns from qualifying purchases, at no extra cost to you. It does not influence our independent rankings.

Strengths

  • Hybrid Mamba-2 architecture cuts long-context memory by ~70%
  • MoE design keeps active params at 9B for fast inference
  • Apache 2.0 with no usage restrictions
  • Built with enterprise governance and provenance in mind
  • Strong throughput on commodity multi-GPU setups

Limitations

  • Requires a recent llama.cpp build for hybrid architecture support
  • Tooling ecosystem still catching up to dense Llama-class models
  • Quality trails frontier 30B+ dense models on hard reasoning

Typical workloads

In our catalog grid, Granite 4.0 H-Small 32B-A9B is filed under Efficient Long Context, Enterprise — the use cases where its size/quality trade-off makes the most sense. Its tags translate to concrete workloads: everyday chat, drafting and summarization.

The 125k-token context window covers long chats and mid-sized documents, though very large retrieval workloads will need chunking. The Apache 2.0 license is permissive, so shipping it inside a commercial product raises no special legal questions.

Architecture & training

Architecture: Hybrid Mamba-2/Transformer (9:1) + granular MoE · 32B/9B active

Training: Granite 4.0 family.

Verdict

The most memory-efficient open MoE for long-context enterprise work — pick it when VRAM, license, and 128k context all matter.

Quick start

ollama run granite4:small-h

Or use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.

Similar models worth comparing

Frequently asked questions

How much VRAM does Granite 4.0 H-Small 32B-A9B need?

At the recommended Q4_K_M quantization, Granite 4.0 H-Small 32B-A9B needs about 19 GB of VRAM. Q8_0 takes 35 GB, and unquantized FP16 weights take 64 GB.

Can Granite 4.0 H-Small 32B-A9B run without a GPU?

Yes — with roughly 32 GB of system RAM it runs CPU-only through llama.cpp or Ollama. Expect a fraction of GPU speed, which is fine for background or batch jobs but slow for interactive chat.

What context window does Granite 4.0 H-Small 32B-A9B support?

Granite 4.0 H-Small 32B-A9B supports a 125k-token context window (128,000 tokens).

Can I use Granite 4.0 H-Small 32B-A9B commercially?

Yes. Granite 4.0 H-Small 32B-A9B is released under Apache 2.0, a permissive open-source license that allows commercial use, modification and redistribution.

How fast is Granite 4.0 H-Small 32B-A9B on consumer hardware?

Our compatibility engine estimates on the order of 30 tokens/sec on a mid-range GPU and up to 75 tokens/sec on high-end cards, assuming the quantization fully fits in VRAM.

Which quantization of Granite 4.0 H-Small 32B-A9B should I download first?

Start with Q4_K_M (19 GB) — the standard size/quality sweet spot. Step up to Q5_K_M or Q8_0 only if you have VRAM headroom. On a 24 GB card you can run up to Q5_K_M.

Tools

Is Granite 4.0 H-Small 32B-A9B the right pick for you?

Compute self-hosted ROI → Back to catalog