Confidential AI Models
Frontier inference without exposing prompts, tools, or memory.
OpenAI-compatible APIs run inside hardware-backed TEEs and return proof of the runtime that handled the request.
Verify private inferenceTEE boundary
Private AI calls prompts, keys, tools, and memory stay inside the runtime. providers route the request without becoming the trust boundary. Proof follows the answer. verify GPU, container, model route, and response. same API shape, hardware-backed receipt, private inference. streaming, tool calls, and agent memory keep their normal developer flow. auditors can inspect evidence without reading the prompt. users get answers plus runtime proof, not another black box.
Private LLM catalog
OpenAI-compatible models with hardware-backed privacy and ACI verification. Keep your SDK flow, change the endpoint, and copy the real call when you need it.
Phala: DeepSeek V4.1 Flash 1017 Uncensored
NewBeta
1.0M context
$0.30/M input
$1.20/M output
$0.006/M cache read
MoonshotAI: Kimi K3.1 Pro
1.0M context
$2.00/M input
$10.00/M output
$0.20/M cache read
DeepSeek: DeepSeek V4.1 Flash 1017
Beta
1.0M context
$0.30/M input
$1.20/M output
$0.006/M cache read
Phala: Qwen3.8 27B Uncensored
1M context
$0.30/M input
$1.50/M output
$0.05/M cache read
Z.ai: GLM 5.3 Flash
Beta
1.0M context
$0.15/M input
$0.50/M output
$0.03/M cache read
Integrate in minutes
Keep your OpenAI-compatible client. Point it at the private endpoint, choose a Phala model slug, and read the proof when the output needs an audit trail. To verify the gateway before each request, use Private AI Proxy or the TypeScript SDK.
selected proof
The OpenAI-compatible endpoint terminates inside the verified gateway boundary.
Coding agents
Private AI Proxy verifies the Phala gateway on your machine and gives agents a local API. Agents get a local token instead of your Phala key. The endpoint serves Chat Completions, Responses, and Messages requests.
AI solution paths
The private model endpoint is the first entry point. The same privacy primitive extends to agents, data workflows, and training.
Serve OpenAI-compatible model calls where prompts, outputs, and customer context need encrypted-in-use protection.
1.0M
$0.30/M input
1.0M
$2.00/M input
1.0M
$0.30/M input
1M
$0.30/M input
1.0M
$1.40/M input
1.0M
$0.15/M input
Run agents with keys, tools, memory, and actions inside a verified runtime instead of a visible automation cloud.
Adapt models on proprietary data while keeping datasets, gradients, checkpoints, and evaluation traces inside the boundary.
private training run
01
sealed
02
running
03
private
04
verified
loss curve
proof attached
attestation.json
Move models to sensitive records and return approved outputs without exposing raw data to the model operator.
source
EHR data
source
Customer records
source
Internal docs
TEE clean room
approved output
Questions
Private LLMs are not just another endpoint. They are a deployment choice between SaaS convenience and self-operated AI infrastructure.
A normal LLM API asks you to trust the provider boundary. Phala runs the model call inside hardware-backed TEEs and can attach runtime proof showing what protected the request.
On-prem gives control, but you operate GPUs, model serving, upgrades, and capacity. Phala keeps the API workflow while adding private execution and verifiable runtime state.
Use the OpenAI-compatible API shape: change the base URL, select a private model slug, and keep your existing SDK or agent framework.
The catalog includes coding, reasoning, general chat, embedding, and decision models from open-weight families such as DeepSeek, Qwen, Z.ai GLM, Moonshot Kimi, Google Gemma, Meta Llama, and OpenAI gpt-oss.
Each response can be checked against an ACI attestation report, signed receipt, response hash, and the upstream.verified session recorded by the gateway.
Providers can cache repeated prompt prefixes automatically. Keep system instructions and reference material at the start. Cached input tokens use the model’s listed cache-read rate, or the normal input rate when none is listed. Check usage.prompt_tokens_details.cached_tokens in the response; cache hits depend on the provider and model.
Use a dedicated stack when you need custom models, reserved GPUs, customer-specific deployments, or a stronger compliance and audit boundary than shared inference.
Start building
Deploy private workloads, verify execution, and scale from models to GPU jobs.