NEAR Protocol is expanding the model catalog available through NEAR AI Cloud, combining confidential inference with access to externally hosted frontier models. The platform is positioning privacy and verifiable execution as core infrastructure for AI workloads, while offering developers a broader range of models through a unified interface.
In an official NEAR Protocol update, the project highlighted support for a growing mix of open-source and frontier models. NEAR AI Cloud’s current product data lists more than 50 models overall, with a subset running confidentially inside Trusted Execution Environments. The distinction matters because not every model receives the same hardware-backed privacy guarantees.
NEAR AI Cloud now supports over 40 open-source and frontier models.
Run DeepSeek V4 Flash, Qwen 3.8, and GLM 5.3 inside sealed hardware with a signed attestation on every request. Claude Opus 5 and GPT-5.5 route through Incognito: the lab sees NEAR AI, not your workload. pic.twitter.com/Niz2cBpWhm
— NEAR Protocol (@NEARProtocol) September 5, 2026
TEEs Provide Hardware-Backed Private Inference
For supported confidential models, NEAR AI uses Intel TDX and NVIDIA confidential-computing infrastructure to isolate inference workloads from infrastructure operators. Requests processed inside these TEEs can receive hardware-backed attestation proving that execution occurred within the expected protected environment.
The architecture is designed to protect both model execution and sensitive user data while providing developers with evidence about the environment processing each request. Attestation turns the execution environment itself into a verifiable component of the service, rather than requiring users to rely exclusively on the cloud provider’s assurances.
NEAR AI Cloud supports open and custom models through this confidential-computing framework, while its wider catalog also includes externally provided frontier systems. The platform is therefore evolving into a multi-model inference layer rather than a service tied to a single AI provider or model family.
Incognito Uses a Different Privacy Model
Frontier models routed through NEAR AI’s Incognito mode operate differently. For example, NEAR AI Cloud states that Claude Opus 5 runs through a partner provider rather than inside a NEAR AI GPU TEE. Incognito hides the individual user’s identity from the model provider by routing requests through a shared NEAR AI credential, but the prompt itself is still processed by that external provider.
That means Incognito should not be equated with hardware-confidential inference. NEAR’s own documentation says these externally hosted models do not provide hardware-level privacy, attestation or cryptographic proof of execution. The system instead offers two distinct trust models: TEE-backed confidential inference and identity-obscuring access to external frontier models.
For developers, the expansion increases model choice while making the execution model an important selection criterion. NEAR AI’s value proposition increasingly rests not simply on how many models it offers, but on giving builders clearer control over the privacy and trust assumptions behind each workload.