Only with CPU offload
Olmo 3.1 32B Instruct at Q4_K_M needs 20.0 GB but only 7.0 GB is addressable, so about 72% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 2.7 tokens per second — usable for batch work, painful for chat.
28% on GPU
8K context
Q4_K_M · 18.1 GB
Apache 2.0
Released 10 Dec 2025
The largest fully open model you can audit end to end. Q4 fits 24 GB, tightly.
The VRAM budget
weights 18.1 GB
Weights 18.1 GB
KV cache @ 8K 1.25 GB
Runtime overhead 0.6 GB
Over budget 13.0 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 31.9 GB | 33.7 GB | — | ~1.3 | −0.1% ppl | 26.7 GB over |
| Q6_K | 24.6 GB | 26.4 GB | — | ~1.8 | −0.4% ppl | 19.4 GB over |
| Q5_K_M | 21.3 GB | 23.1 GB | — | ~2.2 | −0.8% ppl | 16.1 GB over |
| Q4_K_M | 18.1 GB | 20.0 GB | — | ~2.7 | −1.9% ppl | 13.0 GB over |
| Q3_K_M | 14.7 GB | 16.5 GB | — | ~3.6 | −5.4% ppl | 9.5 GB over |
| Q2_K | 12.6 GB | 14.4 GB | — | ~4.6 | −15% ppl | 7.4 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model interleaves sliding-window layers (4096 tokens, 16 of 64 layers global), which is why its cache barely grows with context.
How to run it
$ llama-server \
-hf allenai/Olmo-3.1-32B-Instruct:Q4_K_M \
-c 8192 -ngl 18
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 18.1 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.