Meta · Released April 2025
Llama 4 Scout
What it takes to run Llama 4 Scout locally: memory by quantization, the GPUs that handle it, and how fast it runs.
Updated
- Parameters
- 109B (17B active)
- Architecture
- Mixture of experts
- Context
- 10.5M tokens
- License
- Llama 4 Community License
- Memory at Q4_K_M
- ~70GB
What do you need to run Llama 4 Scout?
Llama 4 Scout needs about 70GB of VRAM at Q4_K_M with 8K context (65GB for the weights plus KV cache and overhead). The smallest GPU that runs it comfortably at Q3_K_M is a 72GB card like the RTX PRO 5000 72GB. Because it is a mixture-of-experts model with 17B active parameters, it generates text much faster than a dense model of the same size, and it tolerates partial CPU offload better.
What to buy
GPUs for Llama 4 Scout
Speeds are estimates at the quantization shown with 8K context. We may earn a commission when you buy through these links. Learn more.
Memory
Llama 4 Scout VRAM requirements by quantization
Lower-bit quantizations shrink the model at some cost to quality. Q4_K_M is the usual sweet spot. Totals include the KV cache for the context length shown plus runtime overhead.
| Quantization | Quality | Weights | Total @ 8K ctx | Total @ 32K ctx |
|---|---|---|---|---|
| Q2_K | Rough | 40 GB | 44 GB | 45 GB |
| Q3_K_M | Usable | 53 GB | 56 GB | 58 GB |
| Q4_K_M | RecommendedDefault | 65 GB | 70 GB | 71 GB |
| Q5_K_M | Very good | 77 GB | 81 GB | 83 GB |
| Q6_K | Near-lossless | 89 GB | 94 GB | 95 GB |
| Q8_0 | Practically lossless | 115 GB | 120 GB | 121 GB |
| BF16 | Full precision | 216 GB | 224 GB | 225 GB |
KV cache note: 48 layers, but three of every four use chunked local attention over 8,192-token chunks, and only the 12 global (NoPE) layers keep the whole context: 2 x 12 x 8 KV heads x 128 dims x 2 bytes = 49,152 bytes per token. The chunked layers add a fixed buffer of about 1.2GB (36 x 8 x 128 x 2 x 2 bytes x 8,192 tokens). Runtimes without chunked-attention support cache all 48 layers, which is 196,608 bytes per token.
Performance
How fast Llama 4 Scout runs
Single-user generation speed on popular GPUs that fit the model entirely in memory, at the best quantization each can hold.
RTX PRO 5000 72GB
Q3_K_M · 72GB
~66tok/s
RTX PRO 6000
Q5_K_M · 96GB
~61tok/s
M5 Ultra 512GB
Q8_0 · 512GB
~21tok/s
M5 Ultra 256GB
Q8_0 · 256GB
~21tok/s
M5 Max 128GB
Q5_K_M · 128GB
~19tok/s
M4 Max 128GB
Q5_K_M · 128GB
~17tok/s
M3 Ultra 512GB
Q8_0 · 512GB
~14tok/s
Ryzen AI Max+ 395
Q5_K_M · 128GB
~9.3tok/s
DGX Spark
Q6_K · 128GB
~8.6tok/s
Ryzen AI Max+ PRO 495
Q8_0 · 192GB
~6.7tok/s
Overview
About Llama 4 Scout
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.
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
Setup
How to run Llama 4 Scout
$ ollama run llama4:scout$ llama-server -hf unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF:Q4_K_M -c 8192 -ngl 99$ mlx_lm.chat --model mlx-community/Llama-4-Scout-17B-16E-Instruct-4bitWeights: meta-llama/Llama-4-Scout-17B-16E-Instruct · GGUF: unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF
Alternatives
Models like Llama 4 Scout
OpenAI
gpt-oss-120b
117B (5.1B active) · 128K ctx
Alibaba Qwen
Qwen3.5 122B-A10B
122B (10B active) · 256K ctx
Z.ai
GLM-4.5-Air
106B (12B active) · 128K ctx
Meta
Llama 3.3 70B Instruct
70.6B · 128K ctx
Frequently asked questions
How much VRAM do you need to run Llama 4 Scout?
About 70GB at Q4_K_M with 8K tokens of context: 65.4GB of weights, 1.6GB of KV cache, and 2.6GB of runtime overhead. Higher-precision quants and longer contexts need more; see the table above.
What is the cheapest GPU that can run Llama 4 Scout?
The smallest card in our database that runs it comfortably is the RTX PRO 5000 72GB (72GB), at Q3_K_M and ~66 tokens/sec.
Is Llama 4 Scout free for commercial use?
Llama 4 Scout is released under the Llama 4 Community License. Commercial use is allowed with some conditions; read the license before shipping a product.
How long a context does Llama 4 Scout support?
Up to 10,485,760 tokens (10.5M). Long contexts need a lot of extra memory for the KV cache, so most local setups run 8K to 32K.
Sources: Llama 4 Scout model card (Hugging Face), modelfit.io GGUF file sizes for Llama 4 Scout. Capability score basis: Placed between the Llama 3.1 8B and Qwen3 32B anchors from its Artificial Analysis Intelligence Index v3 standing (October 2025).