Only with CPU offload

Qwen3.5 397B-A17B at Q4_K_M needs 227.4 GB but only 190.4 GB is addressable, so about 16% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 8.5 tokens per second — usable for batch work, painful for chat.

84% on GPU 8K context Q4_K_M · 226.6 GB Apache 2.0 Released 16 Feb 2026 Vision Not in the Ollama library

Flagship-class at 17B active. A 256 GB Mac Studio or a multi-GPU box at Q4.

Hardware requirements What hardware do I need to run Qwen3.5 397B-A17B? →

At Q4_K_M with an 8K context: a 512 GB Mac (M3 Ultra). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

Fits instead: Qwen3.5 122B-A10B (71.1 GB) · Qwen3.5 9B (6.3 GB)

The VRAM budget

weights 226.6 GB
Weights 226.6 GB KV cache @ 8K 0.23 GB Runtime overhead 0.6 GB Over budget 37.0 GB past 190.4 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
Q8_0 398.8 GB 399.6 GB — ~1.6 −0.1% ppl 209.2 GB over
Q6_K 307.8 GB 308.6 GB — ~2.8 −0.4% ppl 118.2 GB over
Q5_K_M 266.0 GB 266.8 GB — ~4.2 −0.8% ppl 76.4 GB over
Q4_K_M 226.6 GB 227.4 GB — ~8.5 −1.9% ppl 37.0 GB over
Q3_K_M 183.4 GB 184.3 GB 217K 159 −5.4% ppl Long context
Q2_K 157.2 GB 158.0 GB 256K 186 −15% ppl Long context

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 15 of its 60 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/Qwen3.5-397B-A17B:Q4_K_M \
    -c 8192 -ngl 50

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

01Download is 226.6 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.
See all models for this rig Compare with another model

Related models on an Instinct MI300X

ModelTotalTok/s est.
Qwen3.5 122B-A10B
125B MoE · same family
Q4_K_M · 71.1 GB 219 Runs great
Qwen3.5 9B
9.65B · same family
Q4_K_M · 6.3 GB 591 Runs great
Qwen3.5 4B
4.66B · same family
Q4_K_M · 3.5 GB 1224 Runs great
Qwen3.5 2B
2.27B · same family
Q4_K_M · 2.0 GB 2514 Runs great
Ornith 1.5 397B-A17B
397B MoE · similar size
Q4_K_M · 224.1 GB ~9.1 CPU offload
GLM-5.3-Flash 320B-A18B
321B MoE · similar size
Q4_K_M · 181.2 GB 122 Tight fit
DeepSeek V4.1 Flash 552B-A16B
552B MoE · similar size
Q4_K_M · 311.0 GB ~3.9 CPU offload

Every model against the Instinct MI300X →

Qwen3.5 397B-A17B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
227.4 GB — Won't fit
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
227.4 GB ~1.5 CPU offload
M4 · 16 GB
10.7 GB usable · 120 GB/s
227.4 GB — Won't fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
227.4 GB ~1.6 CPU offload
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
227.4 GB ~1.6 CPU offload
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
227.4 GB — Won't fit
M3 Ultra · 512 GB
384.0 GB usable · 800 GB/s
227.4 GB 22 Runs great

All 114 devices, by tier →