Choose a model that follows the instruction.

Selected small-tier candidate: the ungated BF16 Ornith 1.5 9B uncensored derivative. Local requests still use deterministic fake inference until exact-revision qualification passes.

Ornith 1.5 community derivative · experimental evidence

Alignment: abliterated — The publisher reports a targeted residual-writing weight edit and a 9-to-0 refusal change on 64 held-out prompts. Tsubasa has not reproduced that behavior.

Limitation: Local requests use deterministic fake inference; the displayed price is a test fixture, not launch pricing.

Selected medium-tier candidate and default alias target: OrcaRouter's gated FP8 Qwen3.8 27B uncensored derivative. Local requests remain fake until exact-revision qualification passes.

Qwen3.8 community derivative · experimental evidence

Alignment: abliterated — The publisher reports targeted residual-writing edits and reduced refusal rates on its evaluation sets. Those measurements are upstream claims; Tsubasa has not reproduced them.

Limitation: Local requests use deterministic fake inference; the displayed price is a test fixture, not launch pricing.

Selected large-tier identity: the MIT-licensed DeepSeek V4 Flash 0731 abliterated derivative. It remains unavailable while its exact artifact, two-GPU placement, prompt encoding, and current worker behavior are unqualified.

DeepSeek V4 community derivative · unavailable evidence

Alignment: abliterated — The pinned source documents an abliterated transformation and measured refusal behavior; Tsubasa has not reproduced the behavior or general-capability retention claims.

Limitation: This is an unavailable demo record; no local or hosted inference route is enabled.

Selected Ultra Preview identity: the gated Kimi K3 Abliterated V1 derivative. It remains unavailable because the exact checkpoint, gated custom code, storage, and cluster-class placement cannot use the current Serverless architecture.

Kimi K3 community derivative · unavailable evidence

Alignment: abliterated — The publisher reports more than 98 percent removal of measured safeguard directions; Tsubasa has not reviewed the gated transformation artifacts or verified behavior.

Limitation: This is an unavailable preview record; no local or hosted inference route is enabled.