Phala has launched beta access to Kev-4B, an open decision model served through its confidential inference infrastructure. According to Phala’s official Kev-4B launch announcement, the model runs inside a TDX-attested enclave and is available through the /v1/systemone inference endpoint, giving developers a hardware-verifiable environment for workloads that may contain sensitive operational data.
The release differs from a conventional chatbot deployment. Kev-4B takes one shared state and a set of structured questions, then returns probability distributions rather than generated prose. Its purpose is to give AI agents bounded decisions that application code can threshold, log and route, including yes-or-no assessments, choices between predefined options and probability-weighted scores.
Kev-4B Turns Model Outputs Into Typed Decisions
Kev-4B was developed by Jared Palmer as a LoRA adapter and pointer head built on Qwen3.5-4B-Base. Its public model card describes it as an English-language decision model intended for classification, routing, triage and policy-checking workloads. The model computes several independent decisions from the same input state without requiring a text-generation step, making it structurally different from an assistant that produces free-form answers.
Phala’s beta currently validates an 8,192-token context and charges $0.04 per million input tokens, with output priced at $0. The deployment uses model ID jaredpalmer/kev-4b and implements the public System One API. Those economics apply to the current Phala beta and should not be generalized into a broader price comparison with full-scale generative models, since Kev performs a narrower decision workload and does not generate ordinary output tokens.
The model is particularly suited to agent workflows where unbounded language generation may be unnecessary. A support system, for example, can ask whether a case concerns billing, classify its urgency or decide whether confidence is high enough for automatic routing. The application still defines the allowed options and automation thresholds; Kev supplies probability estimates rather than independently determining what actions are permissible.
That distinction is increasingly relevant as confidential AI infrastructure expands beyond conversational inference. NEAR AI has similarly deployed hardware-attested confidential inference for agent workloads, while Bittensor subnets are moving from model benchmarks toward live AI services. The emerging infrastructure layer is separating model capability from the environment in which sensitive inference actually runs.
TDX Attestation Defines the Privacy Boundary
Phala says the Kev-4B deployment exposes receipt, gateway-attestation and attested-session signals that developers can inspect to verify the execution path. Intel TDX isolates the confidential virtual machine from the host environment and provides remote-attestation evidence about that runtime. Attestation can demonstrate that inference occurred inside an expected hardware-isolated environment, but it does not prove that Kev’s decision was accurate or unbiased.
That trust model is different from cryptographic proof systems that mathematically verify a computation. TEEs instead move trust toward processor hardware, firmware and the attestation chain while attempting to prevent infrastructure operators from observing data during execution. The same distinction appears in decentralized GPU infrastructure using confidential computing to protect third-party workloads. Hardware isolation reduces exposure to the cloud operator but does not eliminate every underlying trust assumption.
Phala’s broader private-inference platform supports combined Intel TDX and NVIDIA confidential GPU environments, signed receipts and attested model sessions. However, the Kev-4B launch itself identifies its runtime specifically as a TDX-attested enclave, so NVIDIA Confidential Computing should not be attributed to this model deployment without additional confirmation.
The beta therefore establishes a concrete service rather than evidence of large-scale adoption. Kev-4B is live, priced and accessible through a defined API, but Phala has not disclosed recurring request volume, customer counts or production usage for the model. The measurable development is that a specialized decision model can now be consumed through an attested confidential runtime; whether developers use that architecture repeatedly in production agent systems remains a separate question.