Only with CPU offload
Qwen3.8-Flash-Next 180B-A6B at Q4_K_M needs 102.0 GB but only 7.0 GB is addressable, so about 94% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 4.4 tokens per second — usable for batch work, painful for chat.
The open preview of the Qwen4 architecture: a 125B-A6B hybrid (Gated DeltaNet + sparse attention, KV cache on 12 of 48 blocks) plus a 51B n-gram embedding and a 4B draft head — 180B on disk, 6B active. The hosted "Qwen3.8-Flash" is this model with a 1M window.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 178.1 GB | 178.9 GB | — | ~2.4 | −0.1% ppl | 171.9 GB over |
| Q6_K | 137.5 GB | 138.3 GB | — | ~3.2 | −0.4% ppl | 131.3 GB over |
| Q5_K_M | 118.8 GB | 119.6 GB | — | ~3.7 | −0.8% ppl | 112.6 GB over |
| Q4_K_M | 101.2 GB | 102.0 GB | — | ~4.4 | −1.9% ppl | 95.0 GB over |
| Q3_K_M | 81.9 GB | 82.7 GB | — | ~5.5 | −5.4% ppl | 75.7 GB over |
| Q2_K | 70.2 GB | 71.0 GB | — | ~6.5 | −15% ppl | 64.0 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/Qwen3.8-Flash-Next:Q4_K_M \
-c 8192 -ngl 2
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.