Granite 3 models
9 Granite 3 models from IBM on Hugging Face, from 1.3B to 8.2B parameters, published by IBM in BF16 and F32, with contexts from 4K to 128K tokens. The smallest official model, granite-3.0-1b-a400m-instruct, needs about 3.0 GB of VRAM at its published precision; the cheapest live fit is RTX 3070 at $0.088/hr.
Part of the Granite series · Next generation: Granite 4
Pick a size
One row per official Granite 3 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-3.0-1b-a400m-instruct | 1.3B | 3.0 GB | 1.5 GB | 0.7 GB | RTX 3070 | $0.088/hr | 1.5 GB |
| granite-3.0-1b-a400m-base | 1.4B | 6.2 GB | 1.5 GB | 0.8 GB | RTX 3070 | $0.088/hr | 1.5 GB |
| granite-3.3-2b-instruct | 2.5B | 5.7 GB | 2.8 GB | 1.4 GB | RTX 3070 | $0.088/hr | 2.5 GB |
| granite-guardian-3.3-8b | 8.2B | 18.3 GB | 9.1 GB | 4.6 GB | RTX A5000 | $0.176/hr | 5.0 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 (9)
| Model | Parameters | Published as | VRAM needed | Live GPU fit | Est. $/hr |
|---|---|---|---|---|---|
| granite-3.0-1b-a400m-instruct | 1.3B | BF16 | 3.0 GB | RTX 3070 | $0.088/hr |
| granite-3.0-1b-a400m-base | 1.4B | F32 | 6.2 GB | RTX 3070 | $0.088/hr |
| granite-3.1-2b-instruct | 2.5B | BF16 | 5.7 GB | RTX 3070 | $0.088/hr |
| granite-3.3-2b-instruct | 2.5B | BF16 | 5.7 GB | RTX 3070 | $0.088/hr |
| granite-3.0-8b-instruct | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| granite-3.1-8b-instruct | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| granite-3.2-8b-instruct | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| granite-3.3-8b-instruct | 8.2B | BF16 | 18.3 GB | RTX A5000 | $0.176/hr |
| granite-guardian-3.3-8b | 8.2B | BF16 | 18.3 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 3
- How to run Granite 3: engine support, chat template and context flags, with a serving command sized per model.
- How to fine-tune Granite 3: LoRA and QLoRA memory per model, and the license limits on derivatives.
More from IBM
- Granite 4 (22 models)
- Granite Speech (7 models)
- Granite Docling (1 models)
- Granite Code (1 models)