Only with CPU offload
Granite 4.1 30B at Q4_K_M needs 18.9 GB but only 7.0 GB is addressable, so about 73% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 3.0 tokens per second — usable for batch work, painful for chat.
27% on GPU
8K context
Q4_K_M · 16.3 GB
Apache 2.0
Released 29 Apr 2026
The largest Granite. Dense 29B at Q4 is a comfortable 24 GB fit.
What hardware do I need for Granite 4.1 30B? →
Fits instead: Granite 4.1 8B (6.8 GB) · Granite 4.1 3B (3.1 GB)
The VRAM budget
weights 16.3 GB
Weights 16.3 GB
KV cache @ 8K 2.00 GB
Runtime overhead 0.6 GB
Over budget 11.9 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 28.6 GB | 31.2 GB | — | ~1.5 | −0.1% ppl | 24.2 GB over |
| Q6_K | 22.1 GB | 24.7 GB | — | ~2.0 | −0.4% ppl | 17.7 GB over |
| Q5_K_M | 19.1 GB | 21.7 GB | — | ~2.4 | −0.8% ppl | 14.7 GB over |
| Q4_K_M | 16.3 GB | 18.9 GB | — | ~3.0 | −1.9% ppl | 11.9 GB over |
| Q3_K_M | 13.2 GB | 15.8 GB | — | ~4.0 | −5.4% ppl | 8.8 GB over |
| Q2_K | 11.3 GB | 13.9 GB | — | ~5.0 | −15% ppl | 6.9 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ llama-server \
-hf ibm-granite/granite-4.1-30b:Q4_K_M \
-c 8192 -ngl 17
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 16.3 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.