The local model catalog
The source repository lists the following downloads. The app's download screen shows the exact checkpoint and required space.
| Model | Quantization | Download size | Input |
|---|---|---|---|
| Qwen 3.8 27B | 2-bit S / 4-bit S / 8-bit S | 9.6 / 21.3 / 29.2 GB | Text and images |
| DeepSeek 4 Flash 0731 | 2-bit S | 89.4 GB | Text |
| DeepSeek 4.1 Flash 1017 | 2-bit S / 3-bit S | 376 / 412 GB | Text and images |
DeepSeek 4.1 downloads include a shared 203 GB engram file. Leave additional free space for temporary files, model context caches and project outputs. The figures above describe downloads, not a guarantee that the same amount of free space is enough for installation and ongoing work.
Which model fits your memory
- 16–24 GB: Qwen 3.8 27B runs in 2-bit, all in memory.
- 32–48 GB: Qwen 3.8 27B is recommended in 4-bit, all in memory. DeepSeek 4 Flash and DeepSeek 4.1 Flash 1017 also run here by streaming from your SSD.
- 64 GB: DeepSeek 4 Flash is recommended, streaming from your SSD.
- 96 GB or more: DeepSeek 4.1 Flash 1017 is recommended. DeepSeek 4 Flash runs all in memory from 96 GB, and DeepSeek 4.1 Flash 1017 from 192 GB.
Pick your memory size in the memory grid to see how each model runs.
Memory and storage do different jobs
Unified memory is shared by the model, its context and the rest of your Mac. Storage holds checkpoint files and persistent data. SSD streaming keeps part of a model on your SSD and loads it as needed, so big models run on smaller Macs.
Do not infer that a model runs entirely in memory from total installed RAM alone. Other applications, context length, quantization and the engine's memory policy also matter. A smaller model that responds comfortably can be a better working choice than a larger one that leaves too little headroom.
Try your actual workload
Run a short, representative task first. Check output quality, responsiveness and memory pressure before increasing context or workload. For measured performance, see the official benchmark setup and hardware-specific results.
If local resources are insufficient, a supported remote provider is another option, with different privacy and connectivity tradeoffs. Return to the setup guide to put the choice into practice.