NEAR AI Cloud has added an “abliterated” version of Qwen 3.8 27B to its confidential inference catalog, expanding the platform’s lineup for cybersecurity and red-team workloads. According to NEAR AI’s official model catalog, Qwen 3.8 27B Abliterated accepts text and image inputs and runs inside NEAR AI’s private TEE infrastructure, with pricing set at $0.04 per million input tokens and $1.35 per million output tokens.
NEAR AI announced the addition on November 6, describing it as a model with refusal behavior removed for red teaming and security research. The service lists a context window of roughly 256K tokens and offers trial access without positioning the model as a general consumer chatbot. The release extends a cybersecurity initiative introduced in late October around abliterated versions of Qwen and DeepSeek models.
Abliteration Removes Model Refusals
NEAR AI uses “abliterated” to describe models whose refusal behavior has been reduced or removed, allowing them to answer security questions that conventional assistants may decline. The modification changes model behavior rather than creating a separate cybersecurity guarantee, meaning the model can assist with vulnerability analysis and adversarial testing but does not independently establish that generated findings are correct or safe.
That approach fits a broader expansion of confidential model access across NEAR AI Cloud. The platform now exposes several open-weight models through one inference layer while assigning different privacy tiers depending on how each model is hosted. Qwen 3.8 27B Abliterated belongs to the Private TEE tier, rather than the externally routed model category.
The privacy boundary is hardware-based. NEAR AI says its direct cloud API processes supported models inside Intel TDX virtual machines paired with confidential GPU infrastructure and provides attestation evidence for verification. Intel describes TDX as isolating virtual-machine memory and CPU state from the host and hypervisor. Sensitive prompts can therefore be processed without exposing their contents to ordinary cloud infrastructure outside the trusted environment.
Direct Cloud Access Provides the Privacy Guarantee
That distinction matters because NEAR AI also operates through OpenRouter. Its September integration made NEAR AI available as an OpenRouter provider, but the OpenRouter route does not carry the same confidential-inference guarantee. Requests routed through OpenRouter first pass through infrastructure outside NEAR AI’s attested environment, so NEAR AI recommends its direct API when confidentiality is required.
The new Qwen deployment therefore builds on infrastructure that was already live rather than introducing a new privacy architecture. NEAR AI previously launched confidential inference with hardware-backed attestations, then expanded the catalog and verification tooling around that service. The November addition broadens the workloads that can use the existing private-compute layer, particularly security testing involving sensitive code or undisclosed vulnerabilities.
What remains unproven is adoption depth. NEAR AI has not published request volume, active researcher counts or recurring enterprise usage specifically for Qwen 3.8 27B Abliterated. The measurable development is narrower: a vision-capable, low-refusal Qwen model is now available through NEAR AI’s attested confidential-inference stack at published per-token pricing.