ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Q4_K_M: Smooth Fit 71.92 GB / 94.1 GB Usable

Llama-4-Scout-109B-A17B (MoE) on NVIDIA H20 96GB (HBM3): VRAM Requirement & Speed Estimate

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Llama-4-Scout-109B-A17B (MoE) requires 71.92 GB of VRAM (66.08 GB weights + 3.22 GB KV cache + 2.62 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~241.7 tokens/s.

Concurrency (Batch):
MoE 109B / A17B·48L (d=5120)·40Q/8KV·Ctx 1M·Text+Image
Runnable Models →
GPUs
96 GB VRAM·4000 GB/s·Safe Usable 94.1 GB
Smooth Fit71.92 / 94.1 GB~241.7 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
71.92/ 94.1 GB usable

Physical 96 GB · Load 76.4%

Theoretical Decode Speed
~241.7tokens/s

MoE: reads only 10.31 GB active weights/step

Max Context at Batch=1
126K/ 1M

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 66.08 GBKV Cache (1×16K): 3.22 GBOverhead: 2.62 GB
🚀 Blazing Fast (Instant CoT)~2.1s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run llama4:scout

H20 96G Fits Llama-4-Scout-109B-A17B (MoE) · Cost & TCO

Approx. $14,000 (~$146/GB VRAM) · 400W TDP

📊 All Quantization Levels for Llama-4-Scout-109B-A17B (MoE) on H20 96G

Quantization Total VRAM (16K) Verdict Est. Speed Max Context
FP16 / BF16 226.88 GB Out of VRAM (OOM) — —
FP8 / Q8_0 122.65 GB Partial RAM Offload (Slow) ~13.8 tok/s —
GGUF Q6_K 95.69 GB Partial RAM Offload (Slow) ~109.2 tok/s 8K
GGUF Q5_K_M 83.6 GB Tight Fit (Watch Context) ~193.4 tok/s 68K
Q4_K_M / AWQ 71.92 GB Smooth Fit ~241.7 tok/s 126K
GGUF IQ4_XS 63.59 GB Smooth Fit ~270.6 tok/s 167K
GGUF Q3_K_M 58.86 GB Smooth Fit ~290.3 tok/s 190K
GGUF Q2_K 48.3 GB Smooth Fit ~346.8 tok/s 243K

💡 Recommended Alternative GPUs & Macs for Llama-4-Scout-109B-A17B (MoE)

Frequently Asked Questions

Can NVIDIA H20 96GB (HBM3) run Llama-4-Scout-109B-A17B (MoE) locally?

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Llama-4-Scout-109B-A17B (MoE) requires 71.92 GB of VRAM (66.08 GB weights + 3.22 GB KV cache + 2.62 GB overhead), fitting within H20 96G's 94.1 GB usable memory at ~241.7 tokens/s.

What is the best quantization for Llama-4-Scout-109B-A17B (MoE) on H20 96G?

The highest-quality quantization that fits cleanly on H20 96G is GGUF Q5_K_M (High Balanced Quality), using 83.6 GB VRAM at ~193.4 tokens/s with up to 68K context.

How much VRAM does Llama-4-Scout-109B-A17B (MoE) need in full FP16 precision?

In unquantized FP16/BF16 precision (16K context), Llama-4-Scout-109B-A17B (MoE) requires 226.88 GB VRAM (218 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 109B parameters must sit in memory, while only 17B active parameters are read per token.