Models & providers
Two ways to give Mindestro a mind: download a model that runs entirely on your device, or bring a key from a provider and calls go straight from your browser to them — the app has no server in between. Every AI feature is assigned to one engine in Settings › AI, and the app always names which engine answered.
Built-in local models
Downloaded once, on your click, from Hugging Face (the model's host). They run over WebGPU on your machine: free, private, offline. Desktop browsers only — the weights are too large for phones.
| Model | Download | Cost |
|---|---|---|
| gemma 3 1b | ~0.6 gb | free |
| qwen 3.5 2b | ~1.1 gb | free |
| ministral 3 3b | ~2.0 gb | free |
| phi-4 mini | ~2.2 gb | free |
| qwen 3.5 4b | ~2.3 gb | free |
The Mac app adds Apple Intelligence — Apple's on-device model, nothing to download. Meaning search uses a separate small embedding model (~25 MB), also downloaded on your click.
Bring your own key
Four providers. Pick ONE active provider in Settings › AI › Models; features assigned to the API run on it, billed to your key. Each key is stored only on your device. Prices are USD per million tokens, snapshotted from each provider's own pricing page — the meter in Settings › Usage records what each call actually cost.
Claude
| Model id | Face | In $/Mtok | Out $/Mtok |
|---|---|---|---|
| claude-haiku-4-5 | Claude Haiku 4.5 (fast, cheap) | 1 | 5 |
| claude-sonnet-5 | Claude Sonnet 5 (capable) | 3 | 15 |
| claude-opus-5 | Claude Opus 5 (frontier) | 5 | 25 |
Sees images. Web search available (metered separately, $10 per 1,000 searches).
OpenAI
| Model id | Face | In $/Mtok | Out $/Mtok |
|---|---|---|---|
| gpt-5.6-luna | GPT-5.6 Luna (fast, cheap) | 0.2 | 1.2 |
| gpt-5.6-terra | GPT-5.6 Terra (capable) | 2 | 12 |
Sees images. No web search here.
Grok
| Model id | Face | In $/Mtok | Out $/Mtok |
|---|---|---|---|
| grok-4.3 | Grok 4.3 (capable) | 1.25 | 2.5 |
| grok-4.5 | Grok 4.5 (frontier) | 2 | 6 |
Sees images. No web search here.
DeepSeek
| Model id | Face | In $/Mtok | Out $/Mtok |
|---|---|---|---|
| deepseek-v4-flash | DeepSeek V4 Flash (fast, cheap) | 0.14 | 0.28 |
| deepseek-v4-pro | DeepSeek V4 Pro (capable) | 0.435 | 0.87 |
Text only — cannot see images. No web search here.
The honesty rules
- Capabilities are never silently downgraded: a provider that cannot see images says so when you attach one, instead of answering blind.
- No engine ever falls back to another by itself. A pass assigned to an unavailable engine is visibly unavailable.
- The chat labels every answer with the engine that wrote it.