Llama 4 Scout 17B-16E on Apple M3 Ultra (96 GB)

Can the Apple M3 Ultra (96 GB) run Llama 4 Scout 17B-16E?

Yes, it runs great

Runs great. At Q3_K_M it needs about 61.2 GB of 72.0 GB usable, with an estimated speed of 54.3 tok/s.

Best quant: Q3_K_M
Generation speed: 54.3 tok/s
Prompt processing: 1,036 tok/s
First token (1K prompt): 1.0 s
Largest safe context: 61K

Best app to run it on the Apple M3 Ultra (96 GB)

On macOS

LM Studio

Desktop app · runs on llama.cpp, MLX · use the MLX build

Runs the MLX build, typically 10–20% faster than GGUF for a model with 17B active parameters on Apple silicon.
Point-and-click: search, download and chat in one window.

Open Llama 4 Scout 17B-16E in LM Studio

Ollama One command with Ollama’s own build of this model.
llama.cpp Usually the first to support new model architectures.

How we choose: the files published for this model, your operating system and GPU, whether it fits fully in memory, and published speed tests (MLX vs llama.cpp on Apple silicon, vLLM vs llama.cpp vs Ollama on NVIDIA). Details.

Download Llama 4 Scout 17B-16E

Quickest · Ollama

One command downloads a ready-to-run version and starts a chat.

Get Ollama at ollama.com/download (opens in a new tab) · model page (opens in a new tab)

ollama run llama4:scout

Quantized files · GGUF

Works in LM Studio, llama.cpp, Jan and Ollama. No sign-in needed.

huggingface.co/ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF (opens in a new tab)

ollama run hf.co/ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF

Original weights · Hugging Face

Full-precision files from Meta, for fine-tuning, vLLM, Transformers.

huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct (opens in a new tab)

Free Hugging Face account needed: request access on the model page, and Meta approves it (usually within minutes to a few hours).

Official repo · GitHub

Reference code, inference scripts and docs from Meta. The weights themselves download from Hugging Face.

github.com/meta-llama/llama-models (opens in a new tab)

Meta’s official Llama repo: model cards, prompt formats and download tools.

git clone https://github.com/meta-llama/llama-models

Links checked 3 days ago against Hugging Face, Ollama and GitHub.

How it’s rated

Quality 60Good quality (B) · 60 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 82/100 (quality, fit, speed and memory use) sets the ranking.: How capable the model itself is at Q3_K_M, from its editorial quality tier scaled by the quantization or precision. It’s the same on any hardware that runs this variant. Scale: A+ 85+ · A 70+ · B 55+ · C 40+ · D under 40.
Runs Runs great: How well it runs on the Apple M3 Ultra (96 GB): speed, capped by how well it fits in memory. Runs great: 30+ tokens/s (or near the image, video or decision speed target) · Runs well: 10+ tokens/s · Slow: 3+ tokens/s · Very slow: under 3 · Won’t run: doesn’t fit even with offload. A tight fit or RAM offload caps it at Runs well; heavy offload caps it at Slow.

Match score

To order lists, each result also gets a match score out of 100 from four parts. It favors the most capable model that still runs well.

Model qualityEditorial quality tier of the model, scaled by the quantization or precision used.	27.2 / 45
Memory fitFull marks when the model fits in the share of unified memory the GPU can use; fewer when it goes past that limit.	25.0 / 25
SpeedGeneration speed against a 40 tokens/s target for LLMs, or render time against a target for image and video.	15.0 / 15
Memory useRewards using the memory you have, so a bigger, better variant beats a tiny one that leaves most memory idle.	15.0 / 15
Match score	82.2 / 100

The match score orders results; it is not shown as a grade. Quality and Runs are the two ratings in the tables.

Every quantization on the Apple M3 Ultra (96 GB)

Quant	Memory needed	Available	Speed	Runs	Quality	Download
`Q2_K`	51.4 GB	72.0 GB	66.3 tok/s	Runs great	50Fair quality (C) · 50 of 100: how capable the model is at Q2_K, whatever it runs on. Match score 77/100 (quality, fit, speed and memory use) sets the ranking.	GGUF files (opens in a new tab)
`Q3_K_M`	61.2 GB	72.0 GB	54.3 tok/s	Runs great	60Good quality (B) · 60 of 100: how capable the model is at Q3_K_M, whatever it runs on. Match score 82/100 (quality, fit, speed and memory use) sets the ranking.	GGUF files (opens in a new tab)
`Q4_K_M`	74.2 GB	72.0 GB	37.0 tok/s	Runs wellover GPU memory limit	68Good quality (B) · 68 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 70/100 (quality, fit, speed and memory use) sets the ranking.	GGUF · part 1 of 2 (opens in a new tab)
`Q5_K_M`	85.1 GB	72.0 GB	31.2 tok/s	Slowfar over GPU memory limit	69Good quality (B) · 69 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 63/100 (quality, fit, speed and memory use) sets the ranking.	GGUF files (opens in a new tab)
`Q6_K`	97.0 GB	72.0 GB	—	Won't run	71Great quality (A) · 71 of 100: how capable the model is at Q6_K, whatever it runs on. Match score 9/100 (quality, fit, speed and memory use) sets the ranking.	GGUF files (opens in a new tab)
`Q8_0`	123 GB	72.0 GB	—	Won't run	72Great quality (A) · 72 of 100: how capable the model is at Q8_0, whatever it runs on. Match score 9/100 (quality, fit, speed and memory use) sets the ranking.
`F16`	225 GB	72.0 GB	—	Won't run	72Great quality (A) · 72 of 100: how capable the model is at F16, whatever it runs on. Match score 9/100 (quality, fit, speed and memory use) sets the ranking.	Original weights (opens in a new tab)

Memory figures assume 8K of context. Longer conversations or documents grow the KV cache.

Other hardware that runs it

68Good quality (B) · 68 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 85/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA RTX PRO 6000 Blackwell Q4_K_M · 90.1 tok/s · Runs great
68Good quality (B) · 68 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 83/100 (quality, fit, speed and memory use) sets the ranking. Apple M4 Max (128 GB) Q4_K_M · 30.5 tok/s · Runs great
68Good quality (B) · 68 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 83/100 (quality, fit, speed and memory use) sets the ranking. Apple M2 Ultra (192 GB) Q4_K_M · 42.6 tok/s · Runs great
69Good quality (B) · 69 of 100: how capable the model is at Q5_K_M, whatever it runs on. Match score 82/100 (quality, fit, speed and memory use) sets the ranking. Apple M3 Ultra (256 GB) Q5_K_M · 36.7 tok/s · Runs great
68Good quality (B) · 68 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 81/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA RTX Spark N1X (6144-core GPU) (128 GB) Q4_K_M · 20.2 tok/s · Runs well
68Good quality (B) · 68 of 100: how capable the model is at Q4_K_M, whatever it runs on. Match score 80/100 (quality, fit, speed and memory use) sets the ranking. NVIDIA GB10 (128 GB) Q4_K_M · 18.5 tok/s · Runs well

Quick answers

How much memory does Llama 4 Scout 17B-16E need on the M3 Ultra 96GB?

At Q3_K_M it needs about 61.2 GB. The M3 Ultra 96GB has 72.0 GB usable.

How fast is Llama 4 Scout 17B-16E on the M3 Ultra 96GB?

About 54.3 tok/s at Q3_K_M, an estimate from the hardware’s memory bandwidth and the size of the model.

Which quantization of Llama 4 Scout 17B-16E should I use on the M3 Ultra 96GB?

Q3_K_M. Of the versions that fit, it is the one that runs best; when two run equally well, the check prefers the higher-rated one.

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.