Qwen3-235B-A22B (MoE) — VRAM Requirements, Quantization Matrix & GPU Guide
Qwen3-235B-A22B (MoE) (Qwen, Apache-2.0) is a Mixture-of-Experts (MoE) architecture with 235B resident parameters (22B active per token), 94 layers, and Grouped-Query Attention (GQA) supporting up to 40,960 context tokens.
Interactive Qwen3-235B-A22B (MoE) Memory & Context Simulator
Precision & Quantization VRAM Matrix (8K / 32K / 128K Context)
| Precision / Format | Effective bpw | Weights Only | Total @ 8K ctx | Total @ 32K ctx | Total @ 128K ctx | Recommended Tier (@ 32K) |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 bpw | 437.72 GB | 446.56 GB | 450.96 GB | 468.59 GB | 8× NVIDIA H100 80GB |
| FP8 | 8.50 bpw | 232.54 GB | 238.30 GB | 242.70 GB | 260.33 GB | Mac Studio 512GB |
| Q8_0 | 8.50 bpw | 232.54 GB | 238.30 GB | 242.70 GB | 260.33 GB | Mac Studio 512GB |
| Q6_K | 6.56 bpw | 179.47 GB | 184.43 GB | 188.83 GB | 206.46 GB | Mac Studio 512GB |
| Q5_K_M | 5.69 bpw | 155.66 GB | 160.27 GB | 164.67 GB | 182.30 GB | Mac Studio 512GB |
| Q4_K_M | 4.85 bpw | 132.68 GB | 136.94 GB | 141.35 GB | 158.97 GB | Mac Studio 192GB |
| AWQ INT4 | 4.50 bpw | 123.11 GB | 127.22 GB | 131.63 GB | 149.26 GB | NVIDIA H200 141GB |
| IQ4_XS | 4.25 bpw | 116.27 GB | 120.28 GB | 124.69 GB | 142.31 GB | NVIDIA H200 141GB |
Hardware Compatibility & Decode Speed (Qwen3-235B-A22B (MoE) @ Q4_K_M, 32K Context)
| Hardware Tier | Memory & Bandwidth | Required (32K) | Fit Verdict | Est. Decode Speed |
|---|---|---|---|---|
| RTX 3090 24GB | 24 GB · 936 GB/s | 143.68 GB (599%) | Out of Memory (OOM) | — |
| RTX 4090 24GB | 24 GB · 1008 GB/s | 143.68 GB (599%) | Out of Memory (OOM) | — |
| RTX 5090 32GB | 32 GB · 1792 GB/s | 143.68 GB (449%) | Out of Memory (OOM) | — |
| Apple M-Max 64GB | 64 GB · 546 GB/s | 141.35 GB (221%) | Out of Memory (OOM) | — |
| Mac Studio 128GB | 128 GB · 546 GB/s | 141.35 GB (110%) | Out of Memory (OOM) | — |
| Mac Studio 192GB | 192 GB · 800 GB/s | 141.35 GB (74%) | Tight Fit (>90% Usable) | 24.0 tok/s |
| NVIDIA A100 80GB | 80 GB · 2039 GB/s | 143.68 GB (180%) | Out of Memory (OOM) | — |
| NVIDIA H100 80GB | 80 GB · 3350 GB/s | 143.68 GB (180%) | Out of Memory (OOM) | — |
Verified Transformer Architecture Parameters
- Checkpoint ID
- Qwen/Qwen3-235B-A22B
- Ollama Tag
- qwen3:235b
- License / Release
- Apache-2.0 · 2025-Q2
- Topology & Attention
- Mixture-of-Experts (MoE)
- Attention Mechanism
- Grouped-Query Attention (GQA)
- Total Params (B)
- 235B
- Active Params / Token (B)
- 22B
- Hidden Layers / Hidden Size
- 94 / 4096
- Query Heads / KV Heads
- 64Q / 4KV (d=128)
- MoE Routed Experts
- 8 active / 128 total
- Vocab / Max Position
- 151,936 / 40,960
Quantized formats store per-block scale factors and zero-points alongside packed integer weights. A 4-bit GGUF Q4_K_M checkpoint averages 4.85 effective bits per parameter (not 4.00), while Q8_0 averages 8.50 bpw. For Mixture-of-Experts models, all routed and shared experts remain resident in memory to prevent token-by-token host paging.
Standard Grouped-Query Attention allocates 2 × L × H_kv × d_head × S × B × bytes_kv across layers. DeepSeek Multi-head Latent Attention (MLA) compresses keys and values into L × (d_kv_lora + d_qk_rope) × S × B × bytes_kv, slashing long-context memory by over 90%. Hybrid sliding-window models cap local-attention layers at min(S, window_size) while scaling global layers linearly with S.
Quick-Start vLLM & Ollama Commands for Qwen3-235B-A22B (MoE)
vllm serve Qwen/Qwen3-235B-A22B \
--tensor-parallel-size 1 \
--max-model-len 32768 \
--max-num-seqs 8 \
--gpu-memory-utilization 0.88 \
--kv-cache-dtype fp8 \
--enable-prefix-caching \
--port 8000# Optional macOS unified-memory ceiling override before launching:
# sudo sysctl iogpu.wired_limit_mb=173015
# Save as Modelfile, then run: ollama create qwen3-235b-a22b-custom -f Modelfile
FROM qwen3:235b
PARAMETER num_ctx 32768
PARAMETER num_batch 512
PARAMETER flash_attn true
# Environment variables for ollama serve:
# OLLAMA_NUM_PARALLEL=8 OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 ollama serveCompare Sibling & Similar-Scale Open-Weight Models
Qwen2.5-72B-Instruct
Qwen · 72.7B (GQA)
Qwen3-32B
Qwen · 32.8B (GQA)
QwQ-32B (Reasoning)
Qwen · 32.8B (GQA)
Qwen2.5-Coder-32B-Instruct
Qwen · 32.8B (GQA)
Qwen3-30B-A3B (MoE)
Qwen · 30.5B (GQA)
Qwen3-14B
Qwen · 14.8B (GQA)