Yes, just — 0.7 GB spare
Qwen3.5 9B at Q4_K_M fits your GeForce RTX 4060 entirely on the GPU at 8K context, at an estimated 30 tokens per second. Past 31K 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 · 5.4 GB
Apache 2.0
Released 28 Feb 2026
Vision
The default for 8–12 GB cards in 2026: beats every older 8B on every published benchmark, with vision.
The VRAM budget
weights 5.4 GB
Weights 5.4 GB
KV cache @ 8K 0.25 GB
Runtime overhead 0.6 GB
Free 0.7 GB of 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 18.0 GB | 18.8 GB | — | ~2.8 | Reference | 11.8 GB over |
| Q8_0 | 9.5 GB | 10.4 GB | — | ~7.6 | −0.1% ppl | 3.4 GB over |
| Q6_K | 7.4 GB | 8.2 GB | — | ~14 | −0.4% ppl | 1.2 GB over |
| Q5_K_M | 6.4 GB | 7.2 GB | — | ~23 | −0.8% ppl | 0.2 GB over |
| Q4_K_M | 5.4 GB | 6.3 GB | 31K | 30 | −1.9% ppl | Recommended |
| Q3_K_M | 4.4 GB | 5.2 GB | 64K | 37 | −5.4% ppl | Long context |
| Q2_K | 3.8 GB | 4.6 GB | 84K | 44 | −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. Only 8 of its 32 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ llama-server \
-hf Qwen/Qwen3.5-9B:Q4_K_M \
-c 8192 -ngl 99
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 5.4 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 0.7 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 31K context.