Qwen3.5-35B-A3B (MoE) on NVIDIA L40S 48GB: VRAM Requirement & Speed Estimate
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3.5-35B-A3B (MoE) requires 23.34 GB of VRAM (21.22 GB weights + 0.4 GB KV cache + 1.72 GB overhead), fitting within L40S 48G's 47 GB usable memory at ~308 tokens/s.
Physical 48 GB · Load 49.6%
MoE: reads only 1.82 GB active weights/step
Enable Q8/Q4 KV quantization to scale batch concurrency
OLLAMA_CONTEXT_LENGTH=16384 ollama run qwen3.5:35b-a3bL40S 48G Fits Qwen3.5-35B-A3B (MoE) · Cost & TCO
Approx. $7,500 (~$156/GB VRAM) · 350W TDP
📊 All Quantization Levels for Qwen3.5-35B-A3B (MoE) on L40S 48G
| Quantization | Total VRAM (16K) | Verdict | Est. Speed | Max Context |
|---|---|---|---|---|
| FP16 / BF16 | 73.1 GB | Out of VRAM (OOM) | — | — |
| FP8 / Q8_0 | 39.63 GB | Smooth Fit | ~183.6 tok/s | 256K |
| GGUF Q6_K | 30.98 GB | Smooth Fit | ~233.8 tok/s | 256K |
| GGUF Q5_K_M | 27.09 GB | Smooth Fit | ~266.5 tok/s | 256K |
| Q4_K_M / AWQ | 23.34 GB | Smooth Fit | ~308 tok/s | 256K |
| GGUF IQ4_XS | 20.67 GB | Smooth Fit | ~346.6 tok/s | 256K |
| GGUF Q3_K_M | 19.15 GB | Smooth Fit | ~373.2 tok/s | 256K |
| GGUF Q2_K | 15.76 GB | Smooth Fit | ~420 tok/s | 256K |
💡 Recommended Alternative GPUs & Macs for Qwen3.5-35B-A3B (MoE)
Frequently Asked Questions
Can NVIDIA L40S 48GB run Qwen3.5-35B-A3B (MoE) locally?
Yes. At Q4_K_M 4-bit quantization with a 16K context window, Qwen3.5-35B-A3B (MoE) requires 23.34 GB of VRAM (21.22 GB weights + 0.4 GB KV cache + 1.72 GB overhead), fitting within L40S 48G's 47 GB usable memory at ~308 tokens/s.
What is the best quantization for Qwen3.5-35B-A3B (MoE) on L40S 48G?
The highest-quality quantization that fits cleanly on L40S 48G is FP8 / Q8_0 (99.9% Near-Lossless), using 39.63 GB VRAM at ~183.6 tokens/s with up to 256K context.
How much VRAM does Qwen3.5-35B-A3B (MoE) need in full FP16 precision?
In unquantized FP16/BF16 precision (16K context), Qwen3.5-35B-A3B (MoE) requires 73.1 GB VRAM (70 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 35B parameters must sit in memory, while only 3B active parameters are read per token.