OrcaRouter

AI red-teaming outfit releasing abliterated (refusal-removed) builds of current models, framed as red-team tooling. Their Qwen3.8-27B-Uncensored (Apache-2.0, ~739 likes) strips refusal from the dense Qwen3.8-27B while keeping vision, tool calling, reasoning and the MTP head — GGUF, MLX, and BF16 source weights for further finetuning.

On the doombox: bartowski/orcarouter_Qwen3.8-27B-Uncensored-Q4_K_M.gguf (19 GB, 8 of 59 llama-swap entries) — bartowski’s GGUF quantization of orcarouter’s abliteration.

Context: this is the “naive abliteration” school — direction subtraction on refusal. Gestalt Labs’ SABER is the explicitly capability-preserving counterpoint (they cite 8+ point TruthfulQA drops from this approach); orcarouter’s card is the honest red-team framing of the same operation. Unsloth (the quantizer of most other local GGUFs) also ships abliterated variants but the local collection takes the orcarouter one.

SAQ-2 (2026-10): their quantization family — “more intelligence per GB for everyone, on any hardware” — with fidelity reporting the vault already credits (PPL delta + top-1 agreement + KLD vs BF16; see Local LLM inference). Community result captured 2026-10-03 (clippings/x-2106172339285299340.md): a user pushed OrcaSAQ-2 Cyber 27B to 219K context on 2× RTX 3060s at 68 tok/s (15.7 GB model, correct long-context retrieval) — “the best local 27B result I have measured so far” (HF discussion thread linked in the tweet). The relevance: this is the same 3090-class experiment envelope as HyperQwen/llama-swap setups — 219K context under 16 GB of VRAM changes what “27B locally” means.

Links: Qwen model family (base) · Gestalt Labs (counterpoint) · Local LLM inference (SAQ-2 fidelity reporting) · red-teaming skill material in the Hermes stack covers the usage side