Yes, just — 1.4 GB spare
Qwen3 32B at Q4_K_M fits your GeForce RTX 4090 entirely on the GPU at 8K context, at an estimated 33 tokens per second. Past 13K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully on GPU
8K context
Q4_K_M · 18.4 GB
Apache 2.0
Released Apr 2025
The classic 24 GB target, and still the strongest local translator under 70B. Qwen3.8 27B is smaller and better at everything else.
The VRAM budget
weights 18.4 GB
Weights 18.4 GB
KV cache @ 8K 2.00 GB
Runtime overhead 0.6 GB
Free 1.4 GB of 22.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 32.5 GB | 35.1 GB | — | ~2.6 | −0.1% ppl | 12.7 GB over |
| Q6_K | 25.0 GB | 27.6 GB | — | ~5.7 | −0.4% ppl | 5.2 GB over |
| Q5_K_M | 21.7 GB | 24.3 GB | — | ~12 | −0.8% ppl | 1.9 GB over |
| Q4_K_M | 18.4 GB | 21.0 GB | 13K | 33 | −1.9% ppl | Recommended |
| Q3_K_M | 14.9 GB | 17.5 GB | 27K | 41 | −5.4% ppl | Long context |
| Q2_K | 12.8 GB | 15.4 GB | 36K | 48 | −15% ppl | Long context |
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 Qwen/Qwen3-32B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 18.4 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.
03Only 1.4 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 13K context.