Qwen3.6-35B-A3B (MoE) on 华为昇腾 Ascend 910B 64GB (HBM2e): VRAM Requirement & Speed Estimate
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3.6-35B-A3B (MoE) requires 23.96 GB of VRAM (21.83 GB weights + 0.4 GB KV cache + 1.74 GB overhead), fitting within 昇腾 910B 64G's 61.4 GB usable memory at ~420 tokens/s.
Physical 64 GB · Load 39%
MoE: reads only 1.82 GB active weights/step
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3-6-35b-a3b昇腾 910B 64G Fits Qwen3.6-35B-A3B (MoE) · Cost & TCO
Approx. $13,200 (~$206/GB VRAM) · 310W TDP
📊 All Quantization Levels for Qwen3.6-35B-A3B (MoE) on 昇腾 910B 64G
| Quantization | Total VRAM (16K) | Verdict | Est. Speed | Max Context |
|---|---|---|---|---|
| FP16 / BF16 | 75.14 GB | Partial RAM Offload (Slow) | ~44.5 tok/s | — |
| FP8 / Q8_0 | 40.72 GB | Smooth Fit | ~255 tok/s | 256K |
| GGUF Q6_K | 31.81 GB | Smooth Fit | ~324.7 tok/s | 256K |
| GGUF Q5_K_M | 27.82 GB | Smooth Fit | ~370.1 tok/s | 256K |
| Q4_K_M / AWQ | 23.96 GB | Smooth Fit | ~420 tok/s | 256K |
| GGUF IQ4_XS | 21.21 GB | Smooth Fit | ~420 tok/s | 256K |
| GGUF Q3_K_M | 19.65 GB | Smooth Fit | ~420 tok/s | 256K |
| GGUF Q2_K | 16.16 GB | Smooth Fit | ~420 tok/s | 256K |
💡 Recommended Alternative GPUs & Macs for Qwen3.6-35B-A3B (MoE)
Frequently Asked Questions
Can 华为昇腾 Ascend 910B 64GB (HBM2e) run Qwen3.6-35B-A3B (MoE) locally?
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3.6-35B-A3B (MoE) requires 23.96 GB of VRAM (21.83 GB weights + 0.4 GB KV cache + 1.74 GB overhead), fitting within 昇腾 910B 64G's 61.4 GB usable memory at ~420 tokens/s.
What is the best quantization for Qwen3.6-35B-A3B (MoE) on 昇腾 910B 64G?
The highest-quality quantization that fits cleanly on 昇腾 910B 64G is FP8 / Q8_0 (99.9% Near-Lossless), using 40.72 GB VRAM at ~255 tokens/s with up to 256K context.
How much VRAM does Qwen3.6-35B-A3B (MoE) need in full FP16 precision?
In unquantized FP16/BF16 precision (16K context), Qwen3.6-35B-A3B (MoE) requires 75.14 GB VRAM (72 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 36B parameters must sit in memory, while only 3B active parameters are read per token.