# Llama 4 Scout: hardware requirements
> Llama 4 Scout (Meta) is a 109B (17B active) mixture-of-experts model with a 10.5M-token context window, released under the Llama 4 Community License. It needs about 70GB of memory at Q4_K_M with 8K context.
Llama 4 Scout is Meta's most recent open-weight model: a 109B mixture-of-experts with 17B active parameters, native image input, and a headline 10M-token context window. Meta has not released a newer open model, but at its size Scout is outperformed by gpt-oss-120b and Qwen3.5 122B-A10B on reasoning and coding. It needs about 68GB at Q4_K_M with 8K context, so it suits 96GB to 128GB unified-memory machines (Mac Studio, Strix Halo, DGX Spark), or a 24GB GPU plus 64GB of system RAM with the experts offloaded. The 10M context is mostly theoretical locally; even with chunked attention, 1M tokens of KV cache takes about 50GB.

## Recommended hardware
- **Minimum GPU:** RTX PRO 5000 72GB (72GB), Q3_K_M, ~66 tokens/sec
- **Best value:** RTX PRO 5000 72GB (72GB), Q3_K_M, ~66 tokens/sec
- **Fastest single GPU:** RTX PRO 5000 72GB (72GB), Q3_K_M, ~66 tokens/sec
- **Multi-GPU:** 3× RTX 3090 (72GB), Q4_K_M, ~30 tokens/sec
- **Multi-GPU:** 3× RTX 4090 (72GB), Q4_K_M, ~34 tokens/sec
- **Multi-GPU:** 3× RTX 5090 (96GB), Q4_K_M, ~50 tokens/sec
- **Mac / unified memory:** Ryzen AI Max+ 395 (128GB), Q5_K_M, ~9.3 tokens/sec

## Memory by quantization
| Quant | Weights | Total @ 8K ctx | Total @ 32K ctx |
| --- | --- | --- | --- |
| Q2_K | 40.4GB | 43.8GB | 45.1GB |
| Q3_K_M | 52.6GB | 56.3GB | 57.5GB |
| Q4_K_M | 65.4GB | 69.5GB | 70.7GB |
| Q5_K_M | 76.8GB | 81.3GB | 82.5GB |
| Q6_K | 88.9GB | 93.8GB | 95.0GB |
| Q8_0 | 114.5GB | 120.2GB | 121.4GB |
| BF16 | 215.6GB | 224.3GB | 225.5GB |

## GPUs that run it fully in memory (fastest first)
| GPU | Memory | Quant | Speed | Price |
| --- | --- | --- | --- | --- |
| RTX PRO 5000 72GB | 72GB | Q3_K_M | ~66 tok/s | ~$9,100 |
| RTX PRO 6000 | 96GB | Q5_K_M | ~61 tok/s | ~$16,000 |
| M5 Ultra 512GB | 512GB | Q8_0 | ~21 tok/s | n/a |
| M5 Ultra 256GB | 256GB | Q8_0 | ~21 tok/s | ~$9,499 |
| M5 Max 128GB | 128GB | Q5_K_M | ~19 tok/s | ~$5,100 |
| M4 Max 128GB | 128GB | Q5_K_M | ~17 tok/s | $3,499 MSRP |
| M3 Ultra 512GB | 512GB | Q8_0 | ~14 tok/s | $9,499 MSRP |
| Ryzen AI Max+ 395 | 128GB | Q5_K_M | ~9.3 tok/s | ~$3,300 |
| DGX Spark | 128GB | Q6_K | ~8.6 tok/s | ~$6,950 |
| Ryzen AI Max+ PRO 495 | 192GB | Q8_0 | ~6.7 tok/s | ~$6,800 |

## How to run it
- Ollama: `ollama run llama4:scout`
- llama.cpp: `llama-server -hf unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M -c 8192 -ngl 99`
- MLX (Mac): `mlx_lm.chat --model mlx-community/Llama-4-Scout-17B-16E-Instruct-4bit`
- Weights: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct

## Strengths
- Fast for its size: only 17B parameters are active per token
- Native image understanding
- Slow KV cache growth thanks to chunked attention, so long contexts are cheaper than on dense 70B models
- Good multilingual chat and summarization

## Weaknesses
- Weaker reasoning and coding than gpt-oss-120b or Qwen3.5 122B-A10B at similar memory
- Needs about 68GB at Q4_K_M, out of reach for single consumer GPUs without offload
- No reasoning mode
- Llama 4 license adds use restrictions and a 700M monthly user cap
Full page: https://llamabuilds.ai/models/llama-4-scout · Hardware guide: https://llamabuilds.ai/models/llama-4-scout/requirements

---
Source: https://llamabuilds.ai/models/llama-4-scout · Updated October 14, 2026 · Llama Builds. Speeds marked ~ are estimates from our bandwidth model (https://llamabuilds.ai/methodology); measured results are labeled as such.