Only with CPU offload
Qwen3.5 122B-A10B at Q4_K_M needs 71.1 GB but only 22.4 GB is addressable, so about 69% 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 96–128 GB unified-memory model: 122B of knowledge at 10B-active speed.
What hardware do I need for Qwen3.5 122B-A10B? →
Fits instead: Qwen3.5 9B (6.3 GB) · Qwen3.5 4B (3.5 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 123.7 GB | 124.5 GB | — | ~1.7 | −0.1% ppl | 102.1 GB over |
| Q6_K | 95.5 GB | 96.2 GB | — | ~2.3 | −0.4% ppl | 73.8 GB over |
| Q5_K_M | 82.5 GB | 83.3 GB | — | ~2.8 | −0.8% ppl | 60.9 GB over |
| Q4_K_M | 70.3 GB | 71.1 GB | — | ~3.5 | −1.9% ppl | 48.7 GB over |
| Q3_K_M | 56.9 GB | 57.7 GB | — | ~4.8 | −5.4% ppl | 35.3 GB over |
| Q2_K | 48.7 GB | 49.5 GB | — | ~6.1 | −15% ppl | 27.1 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 12 of its 48 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/Qwen3.5-122B-A10B:Q4_K_M \
-c 8192 -ngl 14
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.