text model · Granite · Windows
Can I run Granite 4.0 H Small on Nvidia GeForce RTX 4080 (16GB)?
No. Granite 4.0 H Small needs ~20.4 GB even at Q4_K_M, but Nvidia GeForce RTX 4080 (16GB) only has ~15 GB usable.
Needs ~20.4 GB even at Q4_K_M, but only ~15 GB is usable.
That figure is at a 4k context and moves about ±15% as context length changes.
The gap is about 5.4 GB: Granite 4.0 H Small needs roughly 20.4 GB at Q4_K_M and Nvidia GeForce RTX 4080 (16GB) leaves only about 15 GB usable for a model. The lightest tracked hardware that runs Granite 4.0 H Small is the Nvidia GeForce RTX 4090 (24GB) at 24 GB. See Granite 4.0 H Small on Nvidia GeForce RTX 4090 (24GB).
- Q4_K_M needed
- ~20.4 GB
- Usable on device
- ~15 GB
- Device memory
- 16 GB
Which quant fits
- Parameters
- 32B (MoE, 9B active)
- Q4_K_M size
- 18.23 GB
- Q8_0 size
- 31.91 GB
- Context
- 128k
- Ollama tag
- granite4:small-h
- Memory
- 16 GB vram
- Usable for weights
- ~15 GB
- Power draw
- ~320 W
- Best runtime
- vLLM (Linux) / Ollama (CUDA)
What you can run instead
Run Granite 4.0 H Small on other hardware
FAQ
Can Nvidia GeForce RTX 4080 (16GB) run Granite 4.0 H Small?
No. Granite 4.0 H Small needs ~20.4 GB even at Q4_K_M, but Nvidia GeForce RTX 4080 (16GB) only has ~15 GB usable.
How much memory does Granite 4.0 H Small need?
Nvidia GeForce RTX 4080 (16GB) does not have enough memory. At Q4_K_M the weights are ~18.23 GB; with KV cache and runtime overhead, budget ~20.4 GB at a 4k context. It is a Mixture-of-Experts model (32B total / 9B active), so all experts must stay in memory; memory tracks total params, not active params.
What is the best tool to run Granite 4.0 H Small on Windows?
LM Studio for a simple setup; Ollama (CUDA) for the most speed. AMD GPUs run via Vulkan/ROCm at roughly half CUDA throughput. NVIDIA is the smooth path on Windows.
Embed this
A live badge for your model card or README, updated as the data is. [](https://localmodel.run/can-i-run/granite-4.0-h-small/nvidia-rtx-4080-16gb) Sources
Weights are measured from GGUF files; KV cache and overhead are computed, so totals can vary ~15% with context and runtime. Any tok/s is a bandwidth estimate. See methodology.