Qwen3-235B-A22B (MoE) on NVIDIA H100 80GB SXM (80GB VRAM)
Even in 4-bit Q4_K_M, Qwen3-235B-A22B (MoE) requires 148.87 GB VRAM (2x NVIDIA H100 80GB SXM GPUs needed for full VRAM execution, or heavy CPU RAM offloading at ~20.1 tok/s).
🎛️ Interactive VRAM & KV-Cache Tuner
📊 Real-Time Memory Footprint
148.87 GB / 80 GB🚀 Ready-to-Run Launch Command
ollama run qwen3:235b # Or with llama.cpp server (8K context, Flash Attention): llama-server -hf Qwen/Qwen3-235B-A22B-GGUF -c 8192 -ngl 99 -fa
Need 148.87 GB+ VRAM to run Qwen3-235B-A22B (MoE)?
Recommended cloud setup for 4-bit: Multi-GPU H100 / H200 Node (or 486.23 GB for full FP16). Spin up a pre-configured vLLM / Ollama instance in 60 seconds:
| Provider | Best For | RTX 4090 (24GB) | A100 (80GB) | H100 (80GB) | Bonus / Trial | Action |
|---|---|---|---|---|---|---|
| RunPod | 1-Click vLLM / Ollama Pods | $0.44 / hr | $1.39 / hr | $2.39 / hr | Instant PyTorch / vLLM Pods + Serverless GPU | Deploy on RunPod → |
| Vast.ai | Cheapest Spot GPU Market | $0.29 / hr | $0.95 / hr | $1.79 / hr | Save 50–70% on RTX 4090 / 5090 / A100 rentals | Check Spot GPUs → |
| Vultr Cloud GPU | Enterprise SLA & Bare Metal | — | $1.79 / hr | $2.69 / hr | $300 Free Credit for New Users | Claim $300 Credit → |
| DigitalOcean GPU Droplets | Managed H100 Droplets | — | — | $2.99 / hr | $200 Free Trial (60 Days) | Get $200 Free → |
All Quantization Formats: Qwen3-235B-A22B (MoE) on NVIDIA H100 80GB SXM
| Quantization Format | Bits/Weight | Weights VRAM | KV Cache (8K) | Total VRAM | Fits on 1x NVIDIA H100 80GB SXM (80GB)? | Est. Speed | GPUs Needed |
|---|---|---|---|---|---|---|---|
| FP16 / BF16 (Unquantized) | 16 bpw | 484.1 GB | 1.47 GB | 486.23 GB | No (Over by 406.2GB) | ~6.1 tok/s | 7x NVIDIA H100 80GB SXM |
| FP8 (W8A8 / E4M3) | 8.5 bpw | 257.18 GB | 1.47 GB | 259.3 GB | No (Over by 179.3GB) | ~11.5 tok/s | 4x NVIDIA H100 80GB SXM |
| GGUF Q8_0 (8-bit) | 8.5 bpw | 257.18 GB | 1.47 GB | 259.3 GB | No (Over by 179.3GB) | ~11.5 tok/s | 4x NVIDIA H100 80GB SXM |
| GGUF Q6_K (6-bit) | 6.56 bpw | 198.48 GB | 1.47 GB | 200.61 GB | No (Over by 120.6GB) | ~14.9 tok/s | 3x NVIDIA H100 80GB SXM |
| GGUF Q5_K_M (5-bit) | 5.69 bpw | 172.16 GB | 1.47 GB | 174.28 GB | No (Over by 94.3GB) | ~17.1 tok/s | 3x NVIDIA H100 80GB SXM |
| GGUF Q4_K_M / AWQ (4-bit) | 4.85 bpw | 146.74 GB | 1.47 GB | 148.87 GB | No (Over by 68.9GB) | ~20.1 tok/s | 2x NVIDIA H100 80GB SXM |
| GGUF IQ4_XS (4.25 bpw) | 4.25 bpw | 128.59 GB | 1.47 GB | 130.72 GB | No (Over by 50.7GB) | ~22.9 tok/s | 2x NVIDIA H100 80GB SXM |
| GGUF Q3_K_M (3-bit) | 3.91 bpw | 118.3 GB | 1.47 GB | 120.43 GB | No (Over by 40.4GB) | ~24.9 tok/s | 2x NVIDIA H100 80GB SXM |
Frequently Asked Questions (Qwen3-235B-A22B (MoE) + NVIDIA H100 80GB SXM)
Can NVIDIA H100 80GB SXM (80GB) run Qwen3-235B-A22B (MoE) locally?
No — A single NVIDIA H100 80GB SXM (80GB) cannot fit Qwen3-235B-A22B (MoE) in VRAM. Even in 4-bit Q4_K_M, Qwen3-235B-A22B (MoE) requires 148.87 GB VRAM (2x NVIDIA H100 80GB SXM GPUs needed for full VRAM execution, or heavy CPU RAM offloading at ~20.1 tok/s).
How is KV Cache calculated for Qwen3-235B-A22B (MoE)?
Qwen3-235B-A22B (MoE) uses MoE-GQA with 94 transformer layers, 64 attention heads, and 4 KV heads (head dimension 128). At 8,192 tokens in FP16, KV cache consumes 1.47 GB. Enabling 8-bit KV cache (-ctk q8_0 -ctv q8_0 in llama.cpp or --kv-cache-dtype fp8 in vLLM) cuts this in half to 0.73 GB.