Granite 4.1 30B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
Largest Granite 4.1 tier — a hybrid Mamba2/attention design (interleaved SSM + attention layers), not MoE despite the 4.0 generation using MoE. ~18GB at FP8 (Q4 likely lower, ~12-15GB, unconfirmed). Context extendable to 512K. Apache 2.0.
Granite 4.1 30B needs about 19 GB of VRAM at FP8 (Q4 est. ~12-15GB, unconfirmed) — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 10.7 GB | ~61 tok/s | Fits | |
| Q3_K_M 3.41 bpw | 13.6 GB | ~49 tok/s | Fits | |
| Q4_K_M 4.83 bpw | 18.9 GB | ~37 tok/s | Fits | |
| Q5_K_M 5.67 bpw | 22.1 GB | ~32 tok/s | Tight | |
| Q6_K 6.56 bpw | 25.4 GB | ~4 tok/s | Offload | |
| Q8_0 8.50 bpw | 32.7 GB | ~3 tok/s | Offload | |
| F16 16.00 bpw | 60.8 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Granite 4.1 30B VRAM calculator →
Get Granite 4.1 30B running
The cheapest catalogued GPU that runs Granite 4.1 30B is the AMD Radeon RX 7900 XTX (24 GB).
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Granite 4.1 30B
Install Ollama, then run:
ollama run granite4.1:30bSpecifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- Parameters
- 30 Billion
- Context window
- 128,000
- Architecture
- Hybrid Mamba2 + Attention (SSM)
- Provider
- IBM
- Licence
- Apache 2.0
- Practical deployment
- Single consumer GPU
- Specified at
- FP8 (Q4 est. ~12-15GB, unconfirmed)
- System RAM
- 32 GB
- Record updated
- 2026-04-29
Commercial use permitted. No usage restrictions beyond attribution.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run Granite 4.1 30B on my GPU?
- Granite 4.1 30B on AMD Radeon RX 9060 XT 8GB
- Granite 4.1 30B on Apple M1 Max
- Granite 4.1 30B on Apple M3 Max
- Granite 4.1 30B on Apple M5 Pro
- Granite 4.1 30B on Intel Arc B570
- Granite 4.1 30B on Intel Arc B580
- Granite 4.1 30B on NVIDIA GeForce RTX 3060 (12GB)
- Granite 4.1 30B on NVIDIA GeForce RTX 3070
- Granite 4.1 30B on NVIDIA GeForce RTX 3070 Ti
- Granite 4.1 30B on NVIDIA GeForce RTX 3080 (10GB)
- Granite 4.1 30B on NVIDIA GeForce RTX 4060
- Granite 4.1 30B on NVIDIA GeForce RTX 4070
- Granite 4.1 30B on NVIDIA GeForce RTX 4070 Super
- Granite 4.1 30B on NVIDIA GeForce RTX 4070 Ti
- Granite 4.1 30B on NVIDIA GeForce RTX 5060
- Granite 4.1 30B on NVIDIA GeForce RTX 5060 Ti 8GB
- Granite 4.1 30B on NVIDIA GeForce RTX 5070
- Granite 4.1 30B on NVIDIA L40S
- Granite 4.1 30B on NVIDIA RTX 6000 Ada Generation
Other IBM Granite 4.1 sizes
Granite 4.1 30B — frequently asked questions
How much VRAM does Granite 4.1 30B need?
About 19 GB at FP8 (Q4 est. ~12-15GB, unconfirmed) — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Granite 4.1 30B run on an RTX 4090 (24 GB)?
Yes. Granite 4.1 30B needs about 19 GB at FP8 (Q4 est. ~12-15GB, unconfirmed), inside a 24 GB card, at an estimated 37 tokens/sec.
How do I run Granite 4.1 30B locally?
Install Ollama and run `ollama run granite4.1:30b`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.
What other sizes does IBM Granite 4.1 come in?
Granite 4.1 3B (3 GB), Granite 4.1 8B (6 GB), Granite 4.1 30B (19 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.