Qwen3.5-122B-A10B (MoE) on NVIDIA H20 96GB (HBM3): VRAM Requirement & Speed Estimate
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3.5-122B-A10B (MoE) requires 77.26 GB of VRAM (73.96 GB weights + 0.52 GB KV cache + 2.78 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~420 tokens/s.
Physical 96 GB · Load 82.1%
MoE: reads only 6.06 GB active weights/step
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3.5:122b-a10bH20 96G Fits Qwen3.5-122B-A10B (MoE) · Cost & TCO
Approx. $14,000 (~$146/GB VRAM) · 400W TDP
📊 All Quantization Levels for Qwen3.5-122B-A10B (MoE) on H20 96G
| Quantization | Total VRAM (16K) | Verdict | Est. Speed | Max Context |
|---|---|---|---|---|
| FP16 / BF16 | 250.7 GB | Out of VRAM (OOM) | — | — |
| FP8 / Q8_0 | 134.04 GB | Out of VRAM (OOM) | — | — |
| GGUF Q6_K | 103.86 GB | Partial RAM Offload (Slow) | ~69.5 tok/s | — |
| GGUF Q5_K_M | 90.33 GB | Tight Fit (Watch Context) | ~359.4 tok/s | 165K |
| Q4_K_M / AWQ | 77.26 GB | Smooth Fit | ~420 tok/s | 256K |
| GGUF IQ4_XS | 67.93 GB | Smooth Fit | ~420 tok/s | 256K |
| GGUF Q3_K_M | 62.64 GB | Smooth Fit | ~420 tok/s | 256K |
| GGUF Q2_K | 50.82 GB | Smooth Fit | ~420 tok/s | 256K |
💡 Recommended Alternative GPUs & Macs for Qwen3.5-122B-A10B (MoE)
Frequently Asked Questions
Can NVIDIA H20 96GB (HBM3) run Qwen3.5-122B-A10B (MoE) locally?
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3.5-122B-A10B (MoE) requires 77.26 GB of VRAM (73.96 GB weights + 0.52 GB KV cache + 2.78 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~420 tokens/s.
What is the best quantization for Qwen3.5-122B-A10B (MoE) on H20 96G?
The highest-quality quantization that fits cleanly on H20 96G is GGUF Q5_K_M (High Balanced Quality), using 90.33 GB VRAM at ~359.4 tokens/s with up to 165K context.
How much VRAM does Qwen3.5-122B-A10B (MoE) need in full FP16 precision?
In unquantized FP16/BF16 precision (16K context), Qwen3.5-122B-A10B (MoE) requires 250.7 GB VRAM (244 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 122B parameters must sit in memory, while only 10B active parameters are read per token.