Qwen · Qwen3 · 235.1B parameters · 22.1B active
Qwen3 235B-A22B VRAM requirements
Qwen3 235B-A22B has 94 layers and uses grouped-query attention (4 KV heads). At Q4_K_M the weights come to 132.8 GB.
Won't fit
135.1 GB of 21.8 GB · 621%Short by 113.4 GB. You can run it with 13 of 94 layers on the RTX 4090 and the rest in system RAM, at roughly 3.23 tok/s — usable for batch work, painful for chat. A smaller quantisation or a shorter context is usually the better trade.
Every quantisation of Qwen3 235B-A22B on a RTX 4090
Highlighted row is the highest quality that still fits at 8K context.
| Quantisation | bpw | Weights | Total | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 437.9 GB | 440.2 GB | Won't fit | — | 0.94 |
| INT8 / W8A8 | 8.50 | 232.3 GB | 234.6 GB | Won't fit | — | 1.79 |
| Q8_0 (GGUF) | 8.50 | 232.3 GB | 234.6 GB | Won't fit | — | 1.79 |
| FP8 (E4M3) | 8.00 | 218.9 GB | 221.2 GB | Won't fit | — | 1.92 |
| Q6_K | 6.56 | 179.5 GB | 181.8 GB | Won't fit | — | 2.36 |
| Q5_K_M | 5.67 | 155.2 GB | 157.5 GB | Won't fit | — | 2.74 |
| Q5_K_S | 5.52 | 151.1 GB | 153.4 GB | Won't fit | — | 2.84 |
| Q4_K_M | 4.85 | 132.8 GB | 135.1 GB | Won't fit | — | 3.23 |
| Q4_K_S | 4.58 | 125.5 GB | 127.8 GB | Won't fit | — | 3.44 |
| Q4_0 | 4.55 | 124.7 GB | 127.0 GB | Won't fit | — | 3.46 |
| AWQ 4-bit | 4.25 | 118.0 GB | 120.3 GB | Won't fit | — | 3.68 |
| GPTQ 4-bit | 4.25 | 118.0 GB | 120.3 GB | Won't fit | — | 3.68 |
| MXFP4 | 4.25 | 118.0 GB | 120.3 GB | Won't fit | — | 3.68 |
| IQ4_XS | 4.25 | 116.5 GB | 118.8 GB | Won't fit | — | 3.72 |
| Q3_K_M | 3.91 | 107.2 GB | 109.5 GB | Won't fit | — | 4.10 |
| IQ3_M | 3.70 | 101.5 GB | 103.8 GB | Won't fit | — | 4.36 |
| IQ3_XXS | 3.06 | 84.1 GB | 86.4 GB | Won't fit | — | 5.34 |
| Q2_K | 2.63 | 72.4 GB | 74.7 GB | Won't fit | — | 6.38 |
| IQ2_XXS | 2.06 | 56.9 GB | 59.2 GB | Won't fit | — | 8.55 |
| IQ1_M | 1.75 | 48.4 GB | 50.7 GB | Won't fit | — | 10.4 |
Qwen3 235B-A22B on each GPU
Q4_K_M weights at 8K context, single card, monitor attached.
| GPU | VRAM | GB/s | Verdict | Max ctx | tok/s |
|---|---|---|---|---|---|
| Mac Studio M3 Ultra 256GB | 256 | 819 | Runs comfortably | 128K | 24.7 |
| Mac Studio M4 Max 128GB | 128 | 546 | Won't fit | — | 7.42 |
| H100 SXM 80GB | 80 | 3350 | Won't fit | — | 6.80 |
| NVIDIA DGX Spark (GB10) | 128 | 273 | Won't fit | — | 6.20 |
| A100 80GB | 80 | 2039 | Won't fit | — | 6.02 |
| Ryzen AI Max+ 395 128GB | 128 | 256 | Won't fit | — | 4.85 |
| RTX A6000 | 48 | 768 | Won't fit | — | 4.05 |
| L40S | 48 | 864 | Won't fit | — | 3.90 |
| RTX 5090 | 32 | 1792 | Won't fit | — | 3.83 |
| Mac Mini M4 Pro 48GB | 48 | 273 | Won't fit | — | 3.70 |
| RTX 5080 | 16 | 960 | Won't fit | — | 3.36 |
| RTX 3090 | 24 | 936 | Won't fit | — | 3.36 |
| RTX 5070 Ti | 16 | 896 | Won't fit | — | 3.36 |
| RTX 5060 Ti 16GB | 16 | 448 | Won't fit | — | 3.34 |
| RTX 5070 | 12 | 672 | Won't fit | — | 3.26 |
| RTX 4090 | 24 | 1008 | Won't fit | — | 3.23 |
| RTX 3060 12GB | 12 | 360 | Won't fit | — | 3.07 |
| Radeon RX 7900 XTX | 24 | 960 | Won't fit | — | 3.07 |
| RTX 3080 10GB | 10 | 760 | Won't fit | — | 3.06 |
| RTX 4080 Super | 16 | 736 | Won't fit | — | 3.05 |
| RTX 4070 Ti Super | 16 | 672 | Won't fit | — | 3.05 |
| RTX 4060 Ti 16GB | 16 | 288 | Won't fit | — | 3.01 |
| RTX 4070 Super | 12 | 504 | Won't fit | — | 2.96 |
| RTX 4070 | 12 | 504 | Won't fit | — | 2.96 |
| Arc B580 | 12 | 456 | Won't fit | — | 2.51 |
Architecture
| Parameters | 235.1B |
| Active per token | 22.1B of 128 experts, top-8 |
| Layers | 94 |
| Hidden size | 4096 |
| Attention heads / KV heads | 64 / 4 |
| Head dimension | 128 |
| Vocabulary | 151,936 |
| Trained context | 128K |
| KV cache per 1K tokens | 0 GB |
| Hugging Face | Qwen/Qwen3-235B-A22B |
The Qwen family
Alibaba's series, and the broadest size ladder available — Qwen3 runs from 0.6B to 32B dense, plus 30B-A3B and 235B-A22B as mixture-of-experts. The 152k vocabulary makes the embedding table a large share of a small model's file. Qwen2.5-Coder is the same architecture trained for code.
Direct answers
See the verdictQwen3 235B-A22B on RTX 4090
See the verdictQwen3 235B-A22B on RTX 3090
See the verdictQwen3 235B-A22B on RTX 5080
See the verdictQwen3 235B-A22B on RTX 5070 Ti
See the verdictQwen3 235B-A22B on RTX 5070
See the verdictQwen3 235B-A22B on RTX 4070 Ti Super
See the verdictQwen3 235B-A22B on RTX 4070 Super
See the verdictQwen3 235B-A22B on RTX 4070
See the verdict