Models bundled with Metroon, registry version 10
This is the catalog inside one specific build, not a live check of what your Mac has installed or what a later update might add. The app can adopt a newer signed registry after this page was generated, so treat this as a record of one build rather than as what any installation currently offers.
Registry-listed models. Availability and assignability are checked in the app. Listed sizes are estimates and context values are registry metadata.
19 curated local models. 35 cloud provider and model line pairs across 4 providers.
Runs on this Mac
Local models
Local models run entirely on-device through MLX. Each one still needs certification, a manifest hash and registry currency before the app assigns it to a Citizen. Listed here is not the same as downloaded, certified or ready to run.
| Model | Parameters | Quantization | Size on disk | Context window | Licence |
|---|---|---|---|---|---|
| Meta-Llama-3.3-70B-Instruct-4bit | 70B | 4bit | 36.97 GB | 131,072 | Llama 3.3 |
| DeepSeek-R1-Distill-Qwen-32B-4bit | 32B | 4bit | 17 GB | 16,384 | MIT |
| GLM-4-32B-0414-4bit | 32B | 4bit | 17.5 GB | 32,768 | MIT |
| MiroThinker-1.7-mini-mlx-4Bit | 31B | 4bit | 17.2 GB | 262,144 | Apache 2.0 |
| MiroThinker-v1.5-30B | 30B | 4bit | 17.5 GB | 262,144 | MIT |
| Qwen3-30B-A3B-4bit | 30B | 4bit | 16 GB | 32,768 | Apache 2.0 |
| Qwen3-30B-A3B-Instruct-2507-4bit | 30B | 4bit | 17.2 GB | 262,144 | Apache 2.0 |
| Mistral-Small-3.2-24B-Instruct-2506-4bit | 24B | 4bit | 13.3 GB | 131,072 | Apache 2.0 |
| Phi-4-reasoning-plus-4bit | 14B | 4bit | 9.16 GB | 32,768 | MIT |
| Qwen3-14B-4bit | 14B | 4bit | 7.75 GB | 32,768 | Apache 2.0 |
| gemma-3-12b-it-4bit | 12B | 4bit | 7.6 GB | 32,768 | Gemma |
| gemma-3-12b-it-qat-4bit | 12B | qat-4bit | 7.6 GB | 32,768 | Gemma |
| DeepSeek-R1-0528-Qwen3-8B-4bit | 8B | 4bit | 4.61 GB | 131,072 | MIT |
| Meta-Llama-3.1-8B-Instruct-4bit | 8B | 4bit | 4.53 GB | 131,072 | Llama 3.1 |
| Qwen3-8B-4bit | 8B | 4bit | 4.5 GB | 32,768 | Apache 2.0 |
| gemma-3-4b-it-4bit | 4B | 4bit | 2.7 GB | 32,768 | Gemma |
| Qwen3-4B-4bit | 4B | 4bit | 2.5 GB | 32,768 | Apache 2.0 |
| Phi-4-mini-instruct-4bit | 3.8B | 4bit | 2.3 GB | 16,384 | MIT |
| Phi-4-mini-reasoning-4bit | 3.8B | 4bit | 2.16 GB | 131,072 | MIT |
Beyond the catalog
Adding your own local models
The models above are the ones Metroon curates and offers directly in the app. You can also add your own. Settings accepts a model identifier and downloads it, so a model Metroon has never seen can still become a Citizen.
This is unsupported and you do it at your own risk. Metroon makes no guarantee that a model outside the curated catalog will work. A model has to be in MLX format and has to be one the app's loader can actually run, so models from the same families as the curated ones above have the best chance of working. Anything you add goes through the same certification the curated models do before the app will assign it to a Citizen, and a model that fails that check stays unavailable rather than being assigned. If something you add will not run, the curated catalog is the supported path.
Calls a provider's API
Cloud providers
These are classification patterns the app matches provider responses against, not a live roster of models your account can reach. Some are overlapping or deprecated patterns kept for compatibility, and a working call still depends on the provider's own availability.
| Provider | Model lines |
|---|---|
| Anthropic | claude-3, claude-fable, claude-haiku, claude-mythos, claude-opus, claude-sonnet |
| gemini-flash, gemini-flash-lite, gemini-pro | |
| OpenAI | gpt-4.1, gpt-4.1-mini, gpt-4.1-nano, gpt-4o, gpt-4o-mini, gpt-5-mini, gpt-5-nano, gpt-5.5, gpt-5.5-pro, gpt-5.6, gpt-5.6-luna, gpt-5.6-sol, gpt-5.6-terra, gpt-5.x, gpt-5.x-pro, o1, o1-pro, o3, o3-mini, o3-pro, o4-mini |
| xAI | grok-3, grok-3-mini, grok-4.1-fast, grok-4.x, unknown |
Cloud Citizens call each provider's API directly, so you bring your own key for any provider you want to use. See Models and keys for how billing and key setup work.
Hardware
Memory
Memory is the real limit on how many local models can run at once. The one line we will commit to: 16 GB of unified memory runs cloud Citizens plus at most one small local model at a time. Above that, the app checks certification and available memory before assigning anything larger, Mac by Mac.
Provenance
Registry version 10 · updated 2026-09-07 · digest 710ee00650fd1097… · source: app repo checkout 8e3caa8, Resources/model-registry.json · no released build yet, this describes the current checkout