Llama 4 · Llama-4 · HF: meta-llama/Llama-4-Scout-17B-16E-Instruct
Llama 4 Scout 109B-A17B — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.10 in / $0.30 out per 1M
Parameters
109B (17B act.)
Attention
GQA (8 KV / 40 Q)
Layers × Hidden
48 × 5120
Max Native Context
1024k tokens
MoE Experts
2 / 16
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 207.09 GiB | 210.04 GiB | 214.54 GiB | Lossless Reference |
| Q8_0 | 8.50 | 110.02 GiB | 112.97 GiB | 117.47 GiB | Near-Lossless |
| Q6_K | 6.57 | 85.04 GiB | 87.99 GiB | 92.49 GiB | High Precision |
| Q5_K_M | 5.69 | 73.65 GiB | 76.60 GiB | 81.10 GiB | Balanced High |
| Q4_K_M | 4.85 | 62.77 GiB | 65.73 GiB | 70.23 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 58.24 GiB | 61.20 GiB | 65.70 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 55.01 GiB | 57.96 GiB | 62.46 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 50.61 GiB | 53.56 GiB | 58.06 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 70.2 / 140.4 GiB | 209.3 tok/s | 0.24s | 406k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 70.2 / 79.4 GiB | 146.1 tok/s | 0.24s | 81k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 70.2 / 79.4 GiB | 86.8 tok/s | 0.73s | 81k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 70.2 / 95.0 GiB | 78.1 tok/s | 0.90s | 164k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 70.0 / 199.7 GiB | 33.2 tok/s | 4.42s | 860k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 70.0 / 399.4 GiB | 33.2 tok/s | 4.42s | 1024k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 70.0 / 149.8 GiB | 32.4 tok/s | 4.88s | 560k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 70.0 / 99.8 GiB | 21.8 tok/s | 6.72s | 259k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 70.0 / 99.8 GiB | 15.8 tok/s | 8.99s | 259k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 70.0 / 96.0 GiB | 9.8 tok/s | 5.00s | 245k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | OOM | 70.2 / 31.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 4090 (24 GB) | OOM | 70.2 / 23.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 70.2 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | OOM | 70.2 / 23.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 70.2 / 15.0 GiB | OOM | — | OOM | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | OOM | 70.2 / 23.0 GiB | OOM | — | OOM | Simulate → |
| Cloud NVIDIA L40S (48 GB) | OOM | 70.2 / 47.4 GiB | OOM | — | OOM | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | OOM | 70.2 / 46.6 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | OOM | 70.0 / 49.9 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 70.2 / 15.0 GiB | OOM | — | OOM | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | OOM | 70.0 / 49.9 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 70.2 / 11.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | OOM | 70.0 / 37.4 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | OOM | 70.0 / 49.9 GiB | OOM | — | OOM | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | OOM | 70.0 / 48.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 70.0 / 10.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (24 GB) | OOM | 70.0 / 18.0 GiB | OOM | — | OOM | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | OOM | 70.0 / 48.0 GiB | OOM | — | OOM | Simulate → |
Interactive Llama 4 Scout 109B-A17B Simulator
109B (17B active) · GQA
1792 GB/s · 250 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 164k
Memory Allocation Footprint 70.23 / 95.00 GiB usable
Weights: 62.77 GiB
KV Cache: 6.00 GiB
Overhead: 1.45 GiB
Headroom: 24.77 GiB
Decode Speed
78.1
tok/s (batch=1)
Prefill Speed
4720
prompt tok/s
TTFT Latency
0.90s
for 4,096 tok
Active Params
17.0B
of 109B total
llama.cpp Server Command
llama-server -hf meta-llama/Llama-4-Scout-17B-16E-Instruct:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 0.75 GiB | 64.63 GiB | 117.0 tok/s | COMFORTABLE |
| 8k | 1.50 GiB | 65.73 GiB | 109.3 tok/s | COMFORTABLE |
| 16k | 3.00 GiB | 67.23 GiB | 96.5 tok/s | COMFORTABLE |
| 32k | 6.00 GiB | 70.23 GiB | 78.1 tok/s | COMFORTABLE |
| 64k | 12.00 GiB | 76.23 GiB | 56.6 tok/s | COMFORTABLE |
| 128k | 24.00 GiB | 88.23 GiB | 36.5 tok/s | TIGHT |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$164.99
56.3h active + power
Cloud GPU/mo
$119.72
@$1.85/hr spot/pod
Serverless API/mo
$9.50
API cheaper
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — Llama 4 Scout 109B-A17B
How much VRAM does Llama 4 Scout 109B-A17B require in Q4_K_M?
Llama 4 Scout 109B-A17B weights occupy 62.77 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 70.23 GiB.
What is the cheapest local hardware that runs Llama 4 Scout 109B-A17B?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs Llama 4 Scout 109B-A17B Q4_K_M at ~78.1 tok/s decode and 0.90s TTFT.
How does Llama 4 Scout 109B-A17B scale KV cache at long context?
Llama 4 Scout 109B-A17B uses GQA with 8 KV heads across 48 layers (128 head_dim).