Only with CPU offload

Qwen3.8-Flash-Next 180B-A6B at Q4_K_M needs 102.0 GB but only 44.8 GB is addressable, so about 57% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 7.0 tokens per second — usable for batch work, painful for chat.

43% on GPU 8K context Q4_K_M · 101.2 GB Qwen Community 1.0 Released 24 Aug 2026 Vision

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.

Hardware requirements What hardware do I need to run Qwen3.8-Flash-Next 180B-A6B? →

At Q4_K_M with an 8K context: a 141 GB GPU (H200 SXM) or a 192 GB Mac (M2 Ultra). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

Fits instead: Qwen3.8 27B (16.7 GB)

The VRAM budget

weights 101.2 GB
Weights 101.2 GB KV cache @ 8K 0.19 GB Runtime overhead 0.6 GB Over budget 57.2 GB past 44.8 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
Q8_0 178.1 GB 178.9 GB — ~3.1 −0.1% ppl 134.1 GB over
Q6_K 137.5 GB 138.3 GB — ~4.4 −0.4% ppl 93.5 GB over
Q5_K_M 118.8 GB 119.6 GB — ~5.4 −0.8% ppl 74.8 GB over
Q4_K_M 101.2 GB 102.0 GB — ~7.0 −1.9% ppl 57.2 GB over
Q3_K_M 81.9 GB 82.7 GB — ~10 −5.4% ppl 37.9 GB over
Q2_K 70.2 GB 71.0 GB — ~14 −15% ppl 26.2 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

terminal
$ llama-server \
    -hf Qwen/Qwen3.8-Flash-Next:Q4_K_M \
    -c 8192 -ngl 20

The engine underneath most of the others. Every knob is exposed. More on llama.cpp.

01Download is 101.2 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02llama.cpp and Ollama spread the layers over both cards on their own. The cards take turns on every token, so expect about one card's speed; vLLM's --tensor-parallel-size 2 runs them in parallel instead.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.
See all models for this rig Compare with another model

Related models on a 2× GeForce RTX 3090

ModelTotalTok/s est.
Qwen3.8 27B
27.8B · same family
Q4_K_M · 16.7 GB 36 Runs great
Qwen3.8 2.4T-A95B
2446B MoE · same family
Q4_K_M · 1376.7 GB — Won't fit
Qwen3 235B-A22B
235B MoE · similar size
Q4_K_M · 134.2 GB ~1.6 CPU offload
Qwen3.5 122B-A10B
125B MoE · similar size
Q4_K_M · 71.1 GB ~6.0 CPU offload
Nemotron 3 Super 120B-A12B
124B MoE · similar size
Q4_K_M · 70.4 GB ~5.1 CPU offload
Ling 3.0 Flash 124B-A5B
124B MoE · similar size
Q4_K_M · 70.4 GB ~12 CPU offload
Mistral Small 4 119B-A6B
119B MoE · similar size
Q4_K_M · 67.7 GB ~9.9 CPU offload

Every model against the 2× GeForce RTX 3090 →

Qwen3.8-Flash-Next 180B-A6B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
102.0 GB ~4.3 CPU offload
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
102.0 GB ~4.5 CPU offload
M4 · 16 GB
10.7 GB usable · 120 GB/s
102.0 GB — Won't fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
102.0 GB ~5.2 CPU offload
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
102.0 GB ~5.2 CPU offload
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
102.0 GB — Won't fit
H200 SXM
139.4 GB usable · 4800 GB/s
102.0 GB 331 Runs great
M2 Ultra · 192 GB
144.0 GB usable · 800 GB/s
102.0 GB 62 Runs great

All 114 devices, by tier →