Ornith-1.5-35B-A3B-Heretic-MTP-APEX-GSQ-RCO-GGUF

모델 소개

Ornith-1.5-35B-A3B (vendor: `ornith-ai`) is a ~35B mixture-of-experts model with ~3B activated parameters per token (≈70 GB in bf16), 262,144-token context, built on top of Qwen3.5/Gemma4 by continued pretraining + reinforcement self-improvement. It is a *reasoning model*: its assistant turns open with a ` … ` chain-of-thought block, and it emits ` ` blocks for function calling. Vendor card: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B The uncensored artifacts on HF were produced by the community with **Heretic** (https://github.com/p-e-w/heretic, v1.4.0 lineage) — the fully automated abliteration tool. Abliteration is a *training-free* modification: it estimates the "refusal direction" in the model's activation space and projects that direction out of selected weight matrices, disabling refusal behaviour without any fine-tuning, then validates against a behavioral test suite and keeps the best pass. Heretic therefore changes a small fraction of weights only; the vendor's reasoning/tool-calling behaviour and general capability are otherwise preserved as far as the trial's KL-bounded search allows. 1. **No Heretic run of our own.** Running Heretic on this 35B model locally (4× V100, CPU-only tooling) is a multi-day job with no guarantee of a good pass. Instead we reused the best public Heretic output: `SC117/Ornith-1.5-35B-A3B-Heretic-MTP-APEX-GGUF`, tier `I-Quality` — verified genuine Heretic 1.4.0 output (Trial 62), dequantised ≈ 5.29 bpw. The vendor's own BF16 repo (`ornith-ai/Ornith-1.5-35B-A3B-GGUF` / `ornith-ai/Ornith-1.5-35B-A3B`) is *not* abliterated, and no BF16 Heretic checkpoint was ever published, so every public uncensor of this model is one Heretic generation downstream of a full-precision master. 2. **The source is already quantized.** The abliterated master was never released at bf16; the best public artifact is a carefully-made quant at ~5.3 bpw. Our builds are therefore **second-generation quants** (Heretic abliteration → 5.29 bpw master → our 3.0/3.5 bpw), which costs some quality versus a hypothetical GSQ-RCO built straight from the vendor's bf16. That is the highest-fidelity uncensored pipeline available for this model as of today.

주요 특징

特点摘要待依据来源补齐，不凭模型名称推断能力。

한계 및 유의 사항

구체적 한계를 판단할 출처가 부족합니다. 실제 성능, 지연, 데이터 정책, 라이선스와 버전 수명을 확인하세요.

출처별 요금 및 비용 분석

요금은 채널, 지역, 설정에 따라 다릅니다. 각 채널 요금은 개별 표시되며 제공사의 통일 공식 요금이 아닙니다.

공개 평가 및 실행 설정

暂无已采集的可核验公开评分。模型未上榜不代表能力较低。

자주 묻는 질문

출처 및 추적 기록

许可证：apache-2.0。模型介绍与第三方评测均以原来源的具体版本和方法为准。