Recommended Models
Pick your RAM size below and we'll show you exactly which file to download — one recommendation, a direct download link, and a plain explanation of why it makes sense for your machine.
Check Apple menu > About This Mac, then choose 8, 16, or 32 GB.
Showing picks for 16 GB RAM.
Recommendations
With 16 GB you can run a noticeably stronger model without taxing the machine. E4B doubles the parameter count from E2B and produces clearly better output.
A 6 GB file that makes a real difference in quality over the 8 GB tier.
Why this makes sense
Or try a different trade-off
Q3_K_M
Same model, lighter compression. About 1 GB smaller and faster to load — a reasonable trade if you want more headroom for other apps.
Download from Hugging FaceQ8_0
Same model, near-lossless precision. Uses about 8 GB — close other apps before loading it on a 16 GB machine. Worth it if you want the highest quality this model can produce.
Download from Hugging FaceSetup
Download the file, add it in Mejl, and set it as your active local model. That is the whole setup.
The Hugging Face model pages can look overwhelming. Ignore the repo structure. You only need the single .gguf file. No mmproj file, no tokenizer, no config — just the one file.
Click the download button above. It goes straight to the .gguf file on Hugging Face — no account needed.
Mejl loads the file by path. If you move or delete it later, Mejl loses track. Your Documents or home folder works fine.
Click Browse and select the .gguf file you downloaded. Mejl will start loading it immediately.
This is the App Store sandbox asking for file-access permission. It only happens once per file location.
Under Active local model, select the file you just added. Mejl will use it from that point on.
Quick Answers
A single file containing the model's weights — everything Mejl needs to run it locally. Think of it like an MP3 for AI: one file, no installation, no dependencies. You download it, point Mejl at it, and it stays on your Mac.
No. A GGUF file is just data — model weights stored as numbers. It contains no executable code and cannot open network connections. Mejl loads the weights and runs inference entirely on your Mac using its own engine. Nothing in the file can reach the internet.
Q4 uses 4-bit compression (smaller file, faster), Q8 uses 8-bit (larger, more precise). For email drafts, the quality difference is small — the file size difference is not.
Q4 means 4-bit compression. K_M is the compression algorithm variant — not a different model. Ignore it; the number before the underscore is what matters.
Good To Know
Yes. Go to Settings › Local Models, browse to a different file, and set it as active. You can have multiple models downloaded and switch between them at any time. Mejl doesn't lock you to your first pick.
No. Mejl only loads the model when you're drafting. When you're reading email or not using the app, the model isn't running and isn't consuming meaningful resources.
Local AI in Mejl requires Apple Silicon — M1 or later. Intel Macs are not supported for local model inference.
That's a sign the model is too big for your setup. On an 8 GB Mac, try a lighter quantization (Q4 instead of Q5). On any Mac, closing other memory-heavy apps before loading the model helps. If it's still slow, step down to the next RAM tier's recommendation.