Model & AI Configuration

The Models page is where you configure which AI model the project uses by default and which model categories are available across the project's interfaces.

Code Model

The Code Model section has a single dropdown: choose the primary model used by the project assistant. This model becomes the default across chat sessions, task execution, and code generation unless overridden in a specific session.

Clicking the dropdown opens a searchable model selector.

Models are grouped by execution type. Selecting a model from the list shows a detail panel on the right with:

• Provider and capability tags (for example, Reasoning, Tool Calling, Attachments, Multimodal)
• Accepted inputs: Images, Documents, Text
• Context window size (for example, 1,000,000 tokens)
• Output token limit (for example, 128,000 tokens)
• The model's identifier string (for example, vercel:anthropic/claude-opus-4.6)

Use this panel to compare models before making a selection. The model chosen here sets the project default; any switch made inside the workspace applies to that session only.

Execution Types

Studio exposes models through two execution environments with different privacy guarantees.

Trusted Execution Environment (TEE) — ORGN Gateway

ORGN Gateway models run inside hardware-isolated Trust Domains (Intel TDX + NVIDIA H100 confidential compute). The CPU and GPU execute inference in an environment that is encrypted and isolated from the host operating system, hypervisor, and infrastructure personnel. Every request produces a cryptographic attestation receipt — hardware-signed evidence verifiable in Scanner.

TEE models run on infrastructure provided by NEAR, Phala Network, and Tinfoil.

These models are also accessible programmatically via the Gateway API and are documented in detail on the Gateway models pages.

Language Models (TEE)

Vision Models (TEE)

Audio Models (TEE)

Embedding Models (TEE)

Zero Data Retention (ZDR) — Vercel & OPENCODE

ZDR models run on Vercel's AI infrastructure and via OPENCODE providers under contractual zero data retention commitments. There is no hardware attestation receipt, the privacy guarantee is policy-enforced rather than hardware-verified — but the catalog is significantly broader, including all major frontier models.

Anthropic (via Vercel)

OpenAI (via Vercel)

Google (via Vercel)

Meta (via Vercel)

Mistral (via Vercel)

Alibaba / Qwen (via Vercel)

DeepSeek (via Vercel)

Moonshot (via Vercel)

ZAI (via Vercel)

Other Providers (via Vercel)

Model Access

The Model Access section controls which categories of models are visible across project-scoped interfaces, including the chat UI and worktree creation.

Two categories can be toggled independently:

• Allow ZDR Models: enables standard zero-data-retention models, including Vercel AI Gateway models and non-TEE OPENCODE providers. Suitable for routine development work.
• Allow TEE Models: enables attested TEE models, including NEAR-, Phala-, and Tinfoil-backed models running on confidential infrastructure. Appropriate for sensitive inference or regulated environments.

At least one category must remain enabled at all times so the project always has an available model. Changes take effect immediately using the Instant Apply button.

TEE vs ZDR at a Glance

For full technical details on execution types, attestation, and using models via the API, see the Gateway models overview.