Llama 4 Scout 17B
MoEToolsVisionPopularMeta · 109B (17B active) · Mixture of Experts
MoE with 16 experts, 17B active params
207.4K
1.2K
2025-04
128K tokens
Use cases
Your hardware
16 GB Reference Desktop
Quantization options
What is quantization?| Quant | Bits | VRAM | Disk | Quality | Status |
|---|---|---|---|---|---|
| Q2_K | 2 | 35.4 GB | 33.3 GB | low | Barely runs · ~1 t/s |
| Q3_K_M | 3 | 49.4 GB | 46.6 GB | moderate | Too heavy · ~1 t/s |
| Q4_K_M | 4 | 56.3 GB | 53.3 GB | good | Too heavy · ~1 t/s |
| Q5_K_M | 5 | 70.3 GB | 66.6 GB | good | Too heavy · ~1 t/s |
| Q6_K | 6 | 84.2 GB | 79.9 GB | excellent | Too heavy · ~1 t/s |
| Q8_0 | 8 | 112.2 GB | 106.6 GB | excellent | Too heavy · ~1 t/s |
| F16 | 16 | 223.8 GB | 213.2 GB | lossless | Too heavy · ~1 t/s |
About this model
Llama 4 Scout 17B is a 109B sparse mixture-of-experts (MoE) model activating 17B active per token from Meta in the Llama family, designed primarily for chat, vision, reasoning workflows. Supports structured tool calling; use Q4_K_M or higher when strict JSON schema adherence matters. Supports up to 128K context tokens; start at 4K–8K with Q8_0 KV cache on tight GPUs.
Hardware planning baselines start at 60.9 GB minimum memory and 101.5 GB recommended memory. At Q4_K_M, weights plus baseline runtime overhead use roughly 56.3 GB of RAM/VRAM and 53.3 GB of disk space. At 109B, Llama 4 Scout 17B belongs firmly in the workstation and multi-GPU tier. Running it comfortably requires 101.5 GB of total memory across dual RTX GPUs, high-memory Apple Silicon (Mac Studio Ultra/Max), or MoE hybrid CPU-RAM offloading—mobile devices and standard 16 GB laptops are too constrained for practical speeds.
Deploy Llama 4 Scout 17B locally via Ollama (`llama4:scout`), Ollama with Flash Attention enabled. Under Llama 4 Community License Agreement, review official weight cards before commercial redistribution.
Highlights
- Local Fit Baseline60.9 GB minimum and 101.5 GB recommended RAM/VRAM.
- Primary QuantQ4_K_M targets 56.3 GB memory and 53.3 GB disk.
- Context & Capabilities128K max context · Sparse MoE · Tool calling.
System requirements by device
Memory planning targets 60.9 GB minimum RAM/VRAM and 101.5 GB recommended memory. The baseline quant is Q4_K_M, estimated at 56.3 GB memory and 53.3 GB disk. Actual runtime usage shifts with context length, KV cache precision, and OS background overhead.
| Device | Notes |
|---|---|
| Mac (Apple silicon) | Requires a 48 GB–192 GB Mac Studio or MacBook Pro Max/Ultra with enough unified wired budget for 56.3 GB. |
| Windows / Linux GPU | Use dual/multi-GPU CUDA splitting (--tensor-split) or MoE CPU expert offloading (-ot) paired with 32–128 GB DDR5 system RAM. |
| Phones & Compact Edge Boards | Not practical on phones (60.9 GB min RAM exceeds Android LMK and iOS Jetsam limits). Pick a 2B–4B model on mobile. |
Start with a short session and leave memory for the operating system. A file fitting on disk does not guarantee that it will load in memory.
Estimated memory by quantization
These are rough estimates, not tested minimum requirements. Check the actual file and runtime documentation; context, offloading and operating-system overhead can change the memory needed.
| Quant Tier | Estimated memory | Disk Size | Quality |
|---|---|---|---|
| Q2_K | 35.4 GB | 33.3 GB | low |
| Q3_K_M | 49.4 GB | 46.6 GB | moderate |
| Q4_K_M | 56.3 GB | 53.3 GB | good |
| Q5_K_M | 70.3 GB | 66.6 GB | good |
| Q6_K | 84.2 GB | 79.9 GB | excellent |
| Q8_0 | 112.2 GB | 106.6 GB | excellent |
| F16 | 223.8 GB | 213.2 GB | lossless |
How to run Llama 4 Scout 17B on your device
Start with a short prompt and the smallest quant that still gives the quality you need. A clean first run is more useful than loading a larger file with no memory headroom.
Run Llama 4 Scout 17B on Multi-GPU or High-RAM Desktop
- Plan for 60.9 GB minimum combined memory and 101.5 GB recommended memory for Q4_K_M.
- Use llama.cpp `-ot ".ffn_.*_exps.=CPU"` or `--tensor-split` across multiple GPUs so attention and KV cache remain in fast VRAM.
- Enable Flash Attention (`-fa on`) and Q8_0 KV cache (`-ctk q8_0 -ctv q8_0`) to prevent context buffers from exceeding GPU headroom.
- Avoid attempting to run workstation-scale weights on smartphones or 8 GB laptops; use a 2B–8B checkpoint on portable devices.
Run Llama 4 Scout 17B on High-Memory Apple Silicon Mac
- Verify you are on a 48 GB–192 GB Mac Studio or MacBook Pro Max/Ultra capable of holding 56.3 GB in wired unified memory.
- Launch via `llama-server` or LM Studio with MLX/Metal enabled and set initial context conservatively to 4K–8K.
- Monitor Activity Monitor Memory Pressure to confirm zero SSD swap thrashing during token generation.
- Step down to a lighter quantization (e.g., Q3_K_M or IQ4_XS) if total memory utilization exceeds 75%.
Llama 4 Scout 17B vs alternatives
| Model | Params | Context | Min RAM | License |
|---|---|---|---|---|
| Llama 4 Scout 17B | 109B (17B active) | 128K | 60.9 GB | Llama 4 Community |
| Llama 3.3 70B | 70B | 128K | 39.1 GB | Llama 3.3 Community |
| Llama 3.2 11B Vision | 11B | 128K | 6.1 GB | Llama 3.2 Community |
| Llama 3.1 8B | 8B | 128K | 4.5 GB | Llama 3.1 Community |
| Llama 3.2 3B | 3B | 128K | 1.7 GB | Llama 3.2 Community |
| TinyLlama 1.1B | 1.1B | 2K | 1 GB | Apache 2.0 |
Use this table to choose the smallest nearby model that still matches the job. A related model with lower RAM can be the better local choice when the current page fits only at the edge of available memory.
Frequently asked questions
- What is Llama 4 Scout 17B?
- Llama 4 Scout 17B is a 109B sparse mixture-of-experts model (17B active active per token) from Meta, supporting up to 128K tokens with 60.9 GB minimum RAM and 101.5 GB recommended RAM.
- How much VRAM or RAM does Llama 4 Scout 17B require?
- At Q4_K_M, Llama 4 Scout 17B occupies roughly 56.3 GB of active memory and 53.3 GB on disk. Add 2–6 GB of spare headroom for KV cache buffers, runtime scratch space, and OS apps (101.5 GB recommended).
- What is the best way to run Llama 4 Scout 17B locally?
- Use `ollama run llama4:scout` for a quick CLI/API setup, or load Q4_K_M in LM Studio or llama.cpp with Flash Attention enabled when you want explicit KV cache and layer control.
- Can Llama 4 Scout 17B run on a consumer laptop or smartphone?
- No smartphone or 8–16 GB laptop can hold Llama 4 Scout 17B cleanly in local memory. It targets multi-GPU NVIDIA workstations, high-RAM Mac Studios, or MoE CPU-RAM hybrid rigs with at least 60.9 GB total memory.
- How does Llama 4 Scout 17B compare to Llama 3.3 70B?
- Llama 3.3 70B (70B, 39.1 GB min RAM) sits closest in size. Compare active parameters, context ceiling (128K vs 128K), and license terms (Llama 4 Community) before switching.