Only with CPU offload
Qwen3.8 27B at Q4_K_M needs 16.7 GB but only 7.0 GB is addressable, so about 62% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 3.5 tokens per second — usable for batch work, painful for chat.
The current default local Qwen: dense 27B, text + image + video, 262K context. Only 16 of its 64 blocks keep a KV cache, so long context is cheap.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 27.5 GB | 28.6 GB | — | ~1.6 | −0.1% ppl | 21.6 GB over |
| Q6_K | 21.2 GB | 22.3 GB | — | ~2.3 | −0.4% ppl | 15.3 GB over |
| Q5_K_M | 18.4 GB | 19.5 GB | — | ~2.8 | −0.8% ppl | 12.5 GB over |
| Q4_K_M | 15.6 GB | 16.7 GB | — | ~3.5 | −1.9% ppl | 9.7 GB over |
| Q3_K_M | 12.7 GB | 13.8 GB | — | ~5.0 | −5.4% ppl | 6.8 GB over |
| Q2_K | 10.8 GB | 11.9 GB | — | ~6.6 | −15% ppl | 4.9 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 16 of its 64 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/Qwen3.8-27B:Q4_K_M \
-c 8192 -ngl 24
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.