Granite 4 models
22 Granite 4 models from IBM on Hugging Face, from 352M to 32.2B parameters, published by IBM in BF16, with contexts from 4K to 128K tokens. The smallest official model, granite-4.0-350m, needs about 0.8 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.
Part of the Granite series ยท Previous generation: Granite 3
Pick a size
One row per official Granite 4 size: the VRAM it needs at each precision, the cheapest GPU that holds it today, and the KV cache for a 32K-token context.
| Model | Parameters | Native VRAM | FP8 VRAM | INT4 VRAM | Live GPU fit (native) | Est. $/hr | KV cache at 32K |
|---|---|---|---|---|---|---|---|
| granite-4.0-350m | 352M | 0.8 GB | 0.4 GB | 0.2 GB | RTX 3070 | $0.088/hr | 0.88 GB |
| granite-4.0-1b-base | 1.6B | 3.6 GB | 1.8 GB | 0.9 GB | RTX 3070 | $0.088/hr | 2.5 GB |
| granite-4.0-1b-speech | 2.3B | 5.2 GB | 2.6 GB | 1.3 GB | RTX 3070 | $0.088/hr | 2.5 GB |
| granite-4.0-h-micro | 3.2B | 7.1 GB | 3.6 GB | 1.8 GB | RTX 3070 | $0.088/hr | 0.25 GB |
| granite-4.1-3b | 3.4B | 7.6 GB | 3.8 GB | 1.9 GB | RTX 3070 | $0.088/hr | 2.5 GB |
| granite-4.2-3b | 3.7B | 8.2 GB | 4.1 GB | 2.0 GB | RTX 3060 | $0.110/hr | 2.5 GB |
| granite-vision-4.1-4b | 4.0B | 8.9 GB | 4.5 GB | 2.2 GB | RTX 3060 | $0.110/hr | 2.5 GB |
| granite-4.0-h-tiny | 6.9B | 15.5 GB | 7.8 GB | 3.9 GB | RTX A4000 | $0.167/hr | 0.25 GB |
| granite-guardian-4.1-8b | 8.4B | 18.7 GB | 9.4 GB | 4.7 GB | RTX A5000 | $0.176/hr | 5.0 GB |
| granite-4.1-8b | 8.8B | 19.7 GB | 9.8 GB | 4.9 GB | RTX A5000 | $0.176/hr | 5.0 GB |
| granite-4.1-30b | 28.9B | 64.5 GB | 32.3 GB | 16.1 GB | A100 | $1.31/hr | 8.0 GB |
| granite-4.2-30b | 29.3B | 65.4 GB | 32.7 GB | 16.4 GB | A100 | $1.31/hr | 8.0 GB |
| granite-4.0-h-small | 32.2B | 72.0 GB | 36.0 GB | 18.0 GB | A100 | $1.31/hr | 0.50 GB |
VRAM is weights times a flat 1.2 overhead; the KV cache is a separate per-model figure at 16-bit, shown where the architecture is published. See the methodology.
Official models (18)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| granite-4.0-350m | 352M | BF16 | 0.8 GB | RTX 3070 | $0.088/hr |
| granite-4.0-1b | 1.6B | BF16 | 3.6 GB | RTX 3070 | $0.088/hr |
| granite-4.0-1b-base | 1.6B | BF16 | 3.6 GB | RTX 3070 | $0.088/hr |
| granite-4.0-1b-speech | 2.3B | BF16 | 5.2 GB | RTX 3070 | $0.088/hr |
| granite-4.0-h-micro | 3.2B | BF16 | 7.1 GB | RTX 3070 | $0.088/hr |
| granite-4.0-micro | 3.4B | BF16 | 7.6 GB | RTX 3070 | $0.088/hr |
| granite-4.1-3b | 3.4B | BF16 | 7.6 GB | RTX 3070 | $0.088/hr |
| granite-4.1-3b-llmon-preview | 3.7B | BF16 | 8.2 GB | RTX 3060 | $0.110/hr |
| granite-4.2-3b | 3.7B | BF16 | 8.2 GB | RTX 3060 | $0.110/hr |
| granite-vision-4.1-4b | 4.0B | BF16 | 8.9 GB | RTX 3060 | $0.110/hr |
| granite-4.0-h-tiny | 6.9B | BF16 | 15.5 GB | RTX A4000 | $0.167/hr |
| granite-guardian-4.1-8b | 8.4B | BF16 | 18.7 GB | RTX A5000 | $0.176/hr |
| granite-4.1-8b | 8.8B | BF16 | 19.7 GB | RTX A5000 | $0.176/hr |
| granite-4.1-8b-llmon-preview | 8.8B | BF16 | 19.7 GB | RTX A5000 | $0.176/hr |
| granite-4.2-8b | 8.8B | BF16 | 19.7 GB | RTX A5000 | $0.176/hr |
| granite-4.1-30b | 28.9B | BF16 | 64.5 GB | A100 | $1.31/hr |
| granite-4.2-30b | 29.3B | BF16 | 65.4 GB | A100 | $1.31/hr |
| granite-4.0-h-small | 32.2B | BF16 | 72.0 GB | A100 | $1.31/hr |
Quantized versions (1)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| granite-4.1-8b-fp8 | 8.8B | F8_E4M3 | 9.8 GB | RTX 4070 | $0.154/hr |
Fine-tunes and community models (3)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| antares-1b | 1.8B | BF16 | 4.1 GB | RTX 3070 | $0.088/hr |
| TwIL-LM3-Pro | 3.7B | BF16 | 8.2 GB | RTX 3060 | $0.110/hr |
| TwIL-8B | 8.8B | BF16 | 19.7 GB | RTX A5000 | $0.176/hr |
VRAM is for the precision the model is published in: the weights times a flat 1.2 overhead, with the KV cache not included. See the methodology. The fit shown is the lowest-priced single GPU type that holds the model at that precision, or the lowest-priced multi-GPU set (up to 8) when none does.
Run and fine-tune Granite 4
- How to run Granite 4: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Granite 4: LoRA and QLoRA memory per model, and the license limits on derivatives.
More from IBM
- Granite Speech (7 models)
- Granite 3 (9 models)
- Granite Docling (1 models)
- Granite Code (1 models)