Can Gemma 4 26B-A4B run on Mac Studio M5 Ultra 96GB?
What this means
Gemma 4 26B-A4B fits on the Mac Studio M5 Ultra · 96GB at Q4_0. We estimate it uses 17.7GB of the conservative 88GB working budget.
Estimated decode speed is 270.8–379.2 tokens/s. A roughly 300-word answer may take around 1 seconds.
See how fast it feels
Using the midpoint of our 270.8–379.2 tokens/s estimate for this demo.
Where the memory goes
| Component | Detail | GB |
|---|---|---|
| Model weights | Q4_0 GGUF (14.4 GB) + mmap overhead | 15.1 |
| KV cache | 4K context window | 1.8 |
| Runtime | macOS inference app + compute buffers | 0.8 |
| Total needed | at Q4_0, 4K context | 17.7 |
| Working budget | 96 GB unified memory − conservative macOS reserve | 88 |
| Headroom | remaining inside the working budget | 70.3 |
Pick your quant
| Quant | Download | Memory | Estimated speed | Verdict |
|---|---|---|---|---|
| Q4_0 ★ | 14.4 GB | 17.7 GB | ~270.8–379.2 tok/s | ✓ Runs great |
Get it running on this Mac
Recommended app: LM Studio. Follow the point-and-click steps below.
Download LM Studio from its official site and drag it to Applications. Open it once and allow macOS to launch it. This route uses the graphical app; you do not need its CLI or local-server features.
Search for google/gemma-4-26B-A4B-it-qat-q4_0-gguf. Use the publisher name shown on this page—or paste its full Hugging Face URL. Similarly named community uploads may contain different files.
Open the GGUF download options and select the row containing Q4_0. The download should be about 14.4 GB. Do not select vision-projector or mmproj helper files for a text-only chat.
Choose the downloaded model in the model selector and click Load if prompted. Keep the default local runtime and start with a 4K (4096-token) context. Close memory-heavy apps for the first load.
Try “Explain why the sky is blue in three sentences.” This page estimates 270.8–379.2 tokens/s; it is not an LM Studio measurement unless marked ✓ Verified.
A second reply while offline confirms that the model is running on this Mac.
This report names an exact Hugging Face repository and GGUF quant. LM Studio lets you search that exact source, choose the matching quant, download it, and chat without using Terminal.
An open-source GGUF alternative with friendly hardware-fit hints.
Its MLX engine is experimental, and support for new model architectures can lag.A trusted Ollama catalog model, or later use with coding tools and a local API.
An Ollama package may use different weights or a different quant from this report.One interface for GGUF, MLX, Ollama, documents, and knowledge workflows.
It exposes more choices than the first-chat path and its free license is for personal use.- Model not listed: paste the exact repository google/gemma-4-26B-A4B-it-qat-q4_0-gguf, not only the model nickname.
- Download stalls: confirm there is enough free storage, reconnect to Wi-Fi, and restart the download inside the app.
- Load fails or the app closes: quit memory-heavy apps. In LM Studio, confirm Q4_0 is the model currently loaded. Otherwise use a smaller model from this Mac's results.
- No reply while offline: make sure the downloaded local model—not a remote or cloud model—is selected in the chat.
Your anonymous feedback helps us prioritize which Mac and model paths to retest.
Other models on this Mac
Gemma 4 26B-A4B on other Mac Studio M5 Ultra configurations
FAQ
Can the Mac Studio M5 Ultra · 96GB run Gemma 4 26B-A4B?
Yes at Q4_0. We estimate about 17.7GB of working memory and 270.8–379.2 tokens/s at 4K context.
Which Gemma 4 26B-A4B quant should I use on this Mac?
Q4_0. It is a 14.4GB download and leaves about 70.3GB inside our conservative working budget.
Which app should I use for Gemma 4 26B-A4B on this Mac?
Start with LM Studio. This page gives the complete point-and-click walkthrough.
Are these speeds measured on a Mac Studio M5 Ultra?
No. The range is a formula estimate based on memory bandwidth, model size, active parameters, cooling, and a 4K context. The app, backend, thermals, and prompt can change real performance.