Trust
Model cards
Cortex offers intelligence levels instead of asking you to pick a model. Each card below shows the model a level uses by default. Levels may be re-pointed to other models; this page shows the configuration as of 2026-09-23. The providers involved are listed on the sub-processors page.
On-device
- Minimum plan
- Free
- Default model
- Ternary-Bonsai-2-27B (prism-ml, GGUF)
- Provider
- None — runs locally
- Where it runs
- On your Mac, through llama.cpp.
- Prompts leave your device?
- No. Prompts are processed by the local model on your Mac and are not sent to Swfte or model providers.
- Retention
- Stays on your Mac. On-device use never spends credits.
- Known limits
- Requires an Apple Silicon Mac.
- Model weights are pinned to a fixed Hugging Face revision (6ed5e12b…) and checksum-verified.
- Connectors, sync and telemetry settings are separate from this choice.
- Like every language model, it can produce confident but incorrect output. Check answers that matter.
Low
- Minimum plan
- Free
- Default model
- gpt-6-luna
- Provider
- OpenAI
- Where it runs
- Swfte's cloud gateway, then OpenAI's API.
- Prompts leave your device?
- Yes. Prompts are sent to Swfte and to OpenAI.
- Retention
- Processed by the provider under its API terms. No zero-data-retention agreement is in place yet (in progress, target <ZDR target date — founder to fill>). Swfte-side retention: <Swfte-side retention period — founder to fill>.
- Known limits
- Uses cloud credits.
- Like every language model, it can produce confident but incorrect output. Check answers that matter.
Medium
- Minimum plan
- Pro
- Default model
- gpt-5.6-terraQuick tasks and translation: gpt-5.6-luna.
- Provider
- OpenAI
- Where it runs
- Swfte's cloud gateway, then OpenAI's API.
- Prompts leave your device?
- Yes. Prompts are sent to Swfte and to OpenAI.
- Retention
- Processed by the provider under its API terms. No zero-data-retention agreement is in place yet (in progress, target <ZDR target date — founder to fill>). Swfte-side retention: <Swfte-side retention period — founder to fill>.
- Known limits
- Uses cloud credits.
- Like every language model, it can produce confident but incorrect output. Check answers that matter.
High
- Minimum plan
- Team
- Default model
- gpt-5.6-sol
- Provider
- OpenAI
- Where it runs
- Swfte's cloud gateway, then OpenAI's API.
- Prompts leave your device?
- Yes. Prompts are sent to Swfte and to OpenAI.
- Retention
- Processed by the provider under its API terms. No zero-data-retention agreement is in place yet (in progress, target <ZDR target date — founder to fill>). Swfte-side retention: <Swfte-side retention period — founder to fill>.
- Known limits
- Uses cloud credits.
- Like every language model, it can produce confident but incorrect output. Check answers that matter.
Ultra
- Minimum plan
- Enterprise
- Default model
- claude-opus-5Translation: claude-sonnet-4-6 (Anthropic). Quick tasks: gpt-5.6-luna (OpenAI).
- Provider
- Anthropic, and OpenAI for quick tasks
- Where it runs
- Swfte's cloud gateway, then Anthropic's or OpenAI's API.
- Prompts leave your device?
- Yes. Prompts are sent to Swfte and to Anthropic or OpenAI.
- Retention
- Processed by the provider under its API terms. No zero-data-retention agreement is in place yet (in progress, target <ZDR target date — founder to fill>). Swfte-side retention: <Swfte-side retention period — founder to fill>.
- Known limits
- One level spans two providers, depending on the task.
- Uses cloud credits.
- Like every language model, it can produce confident but incorrect output. Check answers that matter.
Related: Trust centre · Sub-processors · DPA (draft) · Security