Only with CPU offload
Qwen3 30B-A3B at Q4_K_M needs 18.5 GB but only 7.0 GB is addressable, so about 67% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 11 tokens per second — usable for batch work, painful for chat.
33% on GPU
8K context
Q4_K_M · 17.1 GB
Apache 2.0
Released Apr 2025
The MoE that made "3B active" a category. Qwen3.6 35B-A3B is its direct replacement.
The VRAM budget
weights 17.1 GB
Weights 17.1 GB
KV cache @ 8K 0.75 GB
Runtime overhead 0.6 GB
Over budget 11.5 GB past 7.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 30.2 GB | 31.5 GB | — | ~5.1 | −0.1% ppl | 24.5 GB over |
| Q6_K | 23.3 GB | 24.6 GB | — | ~7.1 | −0.4% ppl | 17.6 GB over |
| Q5_K_M | 20.1 GB | 21.5 GB | — | ~8.5 | −0.8% ppl | 14.5 GB over |
| Q4_K_M | 17.1 GB | 18.5 GB | — | ~11 | −1.9% ppl | 11.5 GB over |
| Q3_K_M | 13.9 GB | 15.2 GB | — | ~15 | −5.4% ppl | 8.2 GB over |
| Q2_K | 11.9 GB | 13.2 GB | — | ~19 | −15% ppl | 6.2 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 Qwen/Qwen3-30B-A3B:Q4_K_M \
-c 8192 -ngl 15
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 17.1 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.