Skip to content

Encrypted inference

API

Explore the OpenAI-compatible endpoint and private transport before creating keys, adding credits, and monitoring usage.

Sign in to unlock the full API console
OverviewReference

Models

IDs are mode-provider/model, always in full. 34 models across 5 providers, each group labelled with its privacy contract; GET /v1/models lists every executable paid ID.

PrivateModeTEE
  • tee-privatemode/gpt-oss-120b128K context
    Fast reasoning for agents, code, and everyday work.Text
  • tee-privatemode/kimi-k2.6256K context
    A heavier model for demanding tasks and image input.Text · images
System OneTEE
  • tee-chunklaya/laya-multilingual1M context
    Yes/no, choice, and score decisions over documents of any length, on Laya.Text
abliteratedPassthrough
  • passthrough-runpod/qwen-abliterated16K context
    Qwen3.6 35B-A3B MoE with refusals removed (abliterated). Plain TLS to a RunPod GPU: not attested. Requests are retained encrypted for 30 days for abuse review.Text
TinfoilTEE
  • tee-tinfoil/gpt-oss-120b128K context
    OpenAI's open-weight reasoning model on Tinfoil.Text
  • tee-tinfoil/glm-5-31M context
    Z.ai's flagship for agentic coding and long-horizon tasks, 1M context.Text
  • tee-tinfoil/kimi-k3256K context
    Moonshot's flagship multimodal model for long-context reasoning and agents.Text · images
  • tee-tinfoil/llama3-3-70b128K context
    Meta's Llama 3.3 70B instruction model.Text
  • tee-tinfoil/gemma4-31b256K context
    Google's Gemma 4 31B multimodal model.Text · images
  • tee-tinfoil/glm-5-3-flash1M context
    Z.ai's fast multimodal model, 1M context. Experimental at Tinfoil: the ID or price may change.Text · images
  • tee-tinfoil/deepseek-v4-1-flash1M context
    DeepSeek's efficient MoE model with image input, 1M context. Experimental at Tinfoil: the ID or price may change.Text · images
VenicePassthroughTEE
  • passthrough-venice/openai-gpt-oss-120b125K context
    OpenAI's open-weight reasoning model on Venice.Text
  • passthrough-venice/z-ai-glm-5-31M context
    Z.ai's flagship for agentic coding and long-horizon tasks, 1M context.Text
  • passthrough-venice/z-ai-glm-5-3-flash1M context
    Z.ai's fast multimodal model, 1M context.Text · images
  • passthrough-venice/kimi-k31M context
    Moonshot's flagship multimodal model for long-context reasoning and agents.Text · images
  • passthrough-venice/kimi-k2-6250K context
    Moonshot's 1T-parameter agentic model with image input, 256K context.Text · images
  • passthrough-venice/deepseek-v4-pro-08131M context
    DeepSeek's 1.6T MoE for advanced reasoning and code, 1M context.Text
  • passthrough-venice/deepseek-v4-1-flash1M context
    DeepSeek's efficient MoE model with image input, 1M context.Text · images
  • passthrough-venice/deepseek-v3.2156K context
    DeepSeek's sparse-attention model with strong reasoning and tool use.Text
  • passthrough-venice/qwen3-coder-480b-a35b-instruct-turbo250K context
    Alibaba's Qwen3 Coder 480B, tuned for fast code generation.Text
  • passthrough-venice/qwen-3-8-27b256K context
    Alibaba's dense 27B vision-language model. Beta at Venice: the ID or price may change.Text · images
  • passthrough-venice/qwen3-235b-a22b-instruct-2507125K context
    Alibaba's 235B MoE instruct model for research and long documents.Text
  • passthrough-venice/llama-3.3-70b125K context
    Meta's Llama 3.3 70B instruction model.Text
  • passthrough-venice/google-gemma-4-31b-it250K context
    Google's Gemma 4 31B multimodal model.Text · images
  • passthrough-venice/mistral-small-3-2-24b-instruct250K context
    Mistral's efficient 24B model with image input.Text · images
  • passthrough-venice/minimax-m27193K context
    MiniMax's agentic model built for multi-agent workflows.Text
  • tee-venice/e2ee-gpt-oss-120b-p125K context
    OpenAI's open-weight reasoning model, in a TDX enclave.Text
  • tee-venice/e2ee-glm-5-3-p1M context
    Z.ai's flagship for agentic coding and long-horizon tasks, 1M context, in a TDX enclave.Text
  • tee-venice/e2ee-glm-5-3-flash1M context
    Z.ai's fast model, 1M context, in a TDX enclave.Text
  • tee-venice/e2ee-glm-5-2-p512K context
    Z.ai's GLM 5.2, 512K context, in a TDX enclave.Text
  • tee-venice/e2ee-kimi-k3-p1M context
    Moonshot's flagship for long-context reasoning and agents, in a TDX enclave.Text
  • tee-venice/e2ee-deepseek-v4-flash1M context
    DeepSeek's efficient MoE model, 1M context, in a TDX enclave.Text
  • tee-venice/e2ee-qwen3-8-27b256K context
    Alibaba's dense 27B model, in a TDX enclave.Text
  • tee-venice/e2ee-qwen3-6-35b-a3b31K context
    Alibaba's fast 35B MoE model, in a TDX enclave.Text

Endpoints

Every route the paid API answers. An OpenAI client pointed at the base URL works unchanged; browser-only E2EE features are intentionally separate.

OpenAPI 3.1
  • POST/v1/chat/completionsChat Completions

    Streaming, tool calling, structured outputs, and image input on models that take it.

  • POST/v1/askAsk

    Typed decisions — yes/no, choice, score — over a document of any length. Not chat: `{state, questions}` in, answers out.

  • GET/v1/modelsList models

    Every model ID the key can call, with its provider, privacy mode, and context window.

  • GET/v1/models/{model}Retrieve model

    One model's entry. The ID contains a slash, so it is accepted raw or URL-encoded.

Not yet

Garlic can only serve what a confidential provider runs, and none of these has a model behind it yet.

  • /v1/embeddingsNo confidential embedding model is on either provider's catalogue.
  • /v1/responsesOpenAI's newer surface. Both providers expose chat completions only.
  • /v1/audio/transcriptionsNo confidential speech model is on either provider's catalogue.
  • /v1/images/generationsNo confidential image backend exists to route to.

One endpoint, three modes

The base URL and API key never change. The mode prefix on the model ID picks the privacy contract; the response echoes it in X-Garlic-Privacy.

  1. tee-provider/model

    Ordinary OpenAI JSON. Garlic's edge attests the provider's enclave and encrypts the request into it. The default when no prefix is given.

  2. passthrough-provider/model

    Plaintext forwarded to the provider over TLS with no attestation. Only for providers that list it, and only when asked for by name.

  3. e2ee-provider/model

    Encrypted on your device before it reaches Garlic. Served to browser chat today; the paid API rejects it until usage can be attributed.