Llama 4.5 Scout — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
Llama 4.5 Scout, reported as a June 2026 mid-cycle refresh of Llama 4 Scout retaining the 10M-token context. UNVERIFIED: the specs here are carried from Llama 4 Scout and no primary source has been read. Secondary sources disagree on whether this release happened at all, so treat every figure as provisional until the model card is checked.
Llama 4.5 Scout needs about 67 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Weights plus overhead only: this model's architecture is unpublished, so no KV cache is added.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 36.6 GB | ~13 tok/s | Offload | |
| Q3_K_M 3.41 bpw | 47.3 GB | ~11 tok/s | Offload | |
| Q4_K_M 4.83 bpw | 66.6 GB | — | Too big | |
| Q5_K_M 5.67 bpw | 78.1 GB | — | Too big | |
| Q6_K 6.56 bpw | 90.2 GB | — | Too big | |
| Q8_0 8.50 bpw | 116.6 GB | — | Too big | |
| F16 16.00 bpw | 218.8 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Llama 4.5 Scout VRAM calculator →
Get Llama 4.5 Scout running
The cheapest catalogued GPU that runs Llama 4.5 Scout is the NVIDIA DGX Spark (128 GB).
or compare on Vast.ai from $0.77/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Llama 4.5 Scout
Install Ollama, then run:
ollama run llama-4-5Specifications
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong. Still unconfirmed: existence, params_b_total, params_b_active, context, releaseDate.
Preview — The model is released, but these specs are thin or rest on a single source. Individual fields may be wrong. Still unconfirmed: existence, params_b_total, params_b_active, context, releaseDate.
- Parameters
- 109 Billion (17B active)
- Context window
- 10,000,000
- Architecture
- Mixture-of-Experts
- Provider
- Meta
- Licence
- Llama Community
- Practical deployment
- Workstation (large unified memory or pro GPU)
- Specified at
- Q4_K_M
- System RAM
- 128 GB
- Record updated
- 2026-06-26
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run Llama 4.5 Scout on my GPU?
- Llama 4.5 on AMD Ryzen AI Max+ 395
- Llama 4.5 on Apple M1 Max
- Llama 4.5 on Apple M1 Pro
- Llama 4.5 on Apple M1 Ultra
- Llama 4.5 on Apple M2 Max
- Llama 4.5 on Apple M2 Pro
- Llama 4.5 on Apple M3 Max
- Llama 4.5 on Apple M3 Pro
- Llama 4.5 on Apple M4
- Llama 4.5 on Apple M4 Max
- Llama 4.5 on Apple M5
- Llama 4.5 on Apple M5 Max
- Llama 4.5 on Apple M5 Pro
- Llama 4.5 on NVIDIA A100 80GB (PCIe)
- Llama 4.5 on NVIDIA DGX Spark
- Llama 4.5 on NVIDIA GeForce RTX 5090
- Llama 4.5 on NVIDIA H100 80GB (PCIe)
- Llama 4.5 on NVIDIA L40S
- Llama 4.5 on NVIDIA RTX 6000 Ada Generation
- Llama 4.5 on NVIDIA RTX PRO 6000 Blackwell
Llama 4.5 Scout — frequently asked questions
How much VRAM does Llama 4.5 Scout need?
About 67 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Llama 4.5 Scout run on an RTX 4090 (24 GB)?
No. Llama 4.5 Scout needs about 67 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Llama 4.5 Scout locally?
Install Ollama and run `ollama run llama-4-5`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.