Only with CPU offload
Olmo 3 7B Instruct at Q4_K_M needs 7.2 GB but only 7.0 GB is addressable, so about 5% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 62 tokens per second — usable for batch work, painful for chat.
95% on GPU
8K context
Q4_K_M · 4.1 GB
Apache 2.0
Released 20 Nov 2025
Fully open — training data and code included. Full multi-head attention, so its cache is 4× a GQA 7B at the same context.
The VRAM budget
weights 4.1 GB
Weights 4.1 GB
KV cache @ 8K 2.50 GB
Runtime overhead 0.6 GB
Over budget 0.2 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 13.6 GB | 16.7 GB | — | ~3.6 | Reference | 9.7 GB over |
| Q8_0 | 7.2 GB | 10.3 GB | — | ~9.8 | −0.1% ppl | 3.3 GB over |
| Q6_K | 5.6 GB | 8.7 GB | 1K | ~18 | −0.4% ppl | 1.7 GB over |
| Q5_K_M | 4.8 GB | 7.9 GB | 3K | ~28 | −0.8% ppl | 0.9 GB over |
| Q4_K_M | 4.1 GB | 7.2 GB | 6K | ~62 | −1.9% ppl | 0.2 GB over |
| Q3_K_M | 3.3 GB | 6.4 GB | 12K | 111 | −5.4% ppl | Fits |
| Q2_K | 2.8 GB | 5.9 GB | 16K | 129 | −15% ppl | Fits |
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, 8 of 32 layers global), which is why its cache barely grows with context.
How to run it
$ llama-server \
-hf allenai/Olmo-3-7B-Instruct:Q4_K_M \
-c 8192 -ngl 30
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 4.1 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.