Only with CPU offload

Qwen3 235B-A22B at Q4_K_M needs 134.2 GB but only 30.4 GB is addressable, so about 79% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 1.4 tokens per second — usable for batch work, painful for chat.

21% on GPU 8K context Q4_K_M · 132.1 GB Apache 2.0 Released Apr 2025

Workstation class. Realistically a 192 GB unified-memory or multi-GPU model.

Hardware requirements What hardware do I need to run Qwen3 235B-A22B? →

At Q4_K_M with an 8K context: a 192 GB GPU (Instinct MI300X) or a 256 GB Mac (M3 Ultra). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.

Fits instead: Qwen3 32B (21.0 GB) · Qwen3 Coder 30B-A3B (18.5 GB)

The VRAM budget

weights 132.1 GB
Weights 132.1 GB KV cache @ 8K 1.47 GB Runtime overhead 0.6 GB Over budget 103.8 GB past 30.4 GB

Quantisation ladder

QuantWeightsTotal @ 8KMax contextTok/sQualityFit
Q8_0 232.5 GB 234.6 GB — ~0.7 −0.1% ppl 204.2 GB over
Q6_K 179.5 GB 181.5 GB — ~1.0 −0.4% ppl 151.1 GB over
Q5_K_M 155.1 GB 157.2 GB — ~1.2 −0.8% ppl 126.8 GB over
Q4_K_M 132.1 GB 134.2 GB — ~1.4 −1.9% ppl 103.8 GB over
Q3_K_M 107.0 GB 109.0 GB — ~1.8 −5.4% ppl 78.6 GB over
Q2_K 91.6 GB 93.7 GB — ~2.3 −15% ppl 63.3 GB over

Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.

How to run it

terminal
$ llama-server \
    -hf Qwen/Qwen3-235B-A22B:Q4_K_M \
    -c 8192 -ngl 20

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

01Download is 132.1 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 RTX 5000 Ada

ModelTotalTok/s est.
Qwen3 32B
32.8B · same family
Q4_K_M · 21.0 GB 19 Runs great
Qwen3 30B-A3B
30.5B MoE · same family
Q4_K_M · 18.5 GB 72 Runs great
Qwen3 Coder 30B-A3B
30.5B MoE · same family
Q4_K_M · 18.5 GB 72 Runs great
Qwen3 14B
14.8B · same family
Q4_K_M · 10.2 GB 42 Runs great
DeepSeek V4 Flash 284B-A13B
284B MoE · similar size
Q4_K_M · 160.7 GB ~2.3 CPU offload
Qwen3.8-Flash-Next 180B-A6B
180B MoE · similar size
Q4_K_M · 102.0 GB ~5.6 CPU offload
GLM-5.3-Flash 320B-A18B
321B MoE · similar size
Q4_K_M · 181.2 GB ~1.6 CPU offload

Every model against the RTX 5000 Ada →

Qwen3 235B-A22B on other hardware

DeviceTotalTok/s est.
GeForce RTX 4060 Laptop
7.0 GB usable · 256 GB/s
134.2 GB ~1.2 CPU offload
GeForce RTX 3060 12 GB
10.6 GB usable · 360 GB/s
134.2 GB ~1.2 CPU offload
M4 · 16 GB
10.7 GB usable · 120 GB/s
134.2 GB — Won't fit
GeForce RTX 4090
22.4 GB usable · 1008 GB/s
134.2 GB ~1.3 CPU offload
Radeon RX 7900 XTX
22.4 GB usable · 960 GB/s
134.2 GB ~1.3 CPU offload
M3 Max · 64 GB
48.0 GB usable · 400 GB/s
134.2 GB — Won't fit
Instinct MI300X
190.4 GB usable · 5300 GB/s
134.2 GB 100 Runs great
M3 Ultra · 256 GB
192.0 GB usable · 800 GB/s
134.2 GB 17 Runs great

All 114 devices, by tier →