Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16

An uncensored / abliterated build of deepreinforce-ai/Ornith-1.0-35B — DeepReinforce's state-of-the-art agentic-coding MoE — with refusal behavior removed while preserving capability.

Lineage note: the upstream card says the family is "post-trained on Gemma 4 and Qwen 3.5", but weight-correlation testing shows the 35B-MoE member is initialized from Qwen/Qwen3.6-35B-A3B (rel-L2 0.0022 vs 0.0256 to Qwen3.5-35B-A3B-Base). Architecture: qwen3_5_moe — 40 layers (30 GatedDeltaNet linear-attention + 10 full-attention), 256 routed experts + 1 shared (A3B), vision tower, 256K context.

What was done

A measured, validate-before-ship abliteration:

1. SSM conv1d outlier repair (FernflowerAI method) — rescaled 2 outlier blocks (layers 36/37, σ 0.10→0.062) before abliteration to prevent coherence collapse.

2. Abliteration with abliterix v1.9: grimjim norm-preserving biprojected abliteration + Expert-Granular Abliteration (EGA) across all 256 fused experts + shared expert + router suppression, Optuna multi-objective search (refusals vs KL). Q/K/V left untouched (attn_output_gate). GatedDeltaNet/SSM internals and the vision tower are not modified.

3. Gentle-knee selection. The lowest-refusal trial was over-abliterated (coherent content → word-salad on real generation — the classic "lowest-refusal ≠ shippable" trap). The shipped winner uses a 170× lighter expert edit for the same refusal removal, verified coherent on long generation.

Validation (measured on this model)

| Metric | Original Ornith-1.0-35B | This model |

|---|---|---|

| Refusals (80 diverse harmful prompts: CBRN, cyber, weapons, self-harm) | high | 0 / 80 (0.0%) |

| Agentic/coding pass@1 (18-task self-contained probe) | 0.833 | 0.833 (identical, family-by-family) |

| First-token KL vs base (abliteration fidelity) | — | ~0.0014 |

| Coherence (benign + harmful, long gen) | — | clean (no degeneration) |

Zero coding-capability degradation (the model's core competency — base Ornith scores Terminal-Bench 2.1 = 64.2, SWE-bench Verified = 75.6) and full refusal removal. This is a near-lossless uncensoring: the abliteration is gentle enough (KL ~0.0014) that capability is preserved, while refusals are eliminated.

Quickstart (vLLM)

vllm serve AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16 \
  --served-model-name ornith --max-model-len 262144 \
  --gpu-memory-utilization 0.85 --max-num-batched-tokens 16384 \
  --mamba-cache-dtype float32 --reasoning-parser qwen3 \
  --enable-prefix-caching --trust-remote-code

Reasoning model: every turn opens …. Recommended sampling: temperature 0.6, top_p 0.95, top_k 20. Vision (image/video) is inherited from the base and intact; on a vision-enabled deploy, KV cache stays BF16.

Variants

• -BF16 (this repo) — full precision, source of truth.
• -FP8 — compressed-tensors FP8 for efficient vLLM serving (GatedDeltaNet/router/gates/embeds/vision kept BF16). (coming — vLLM-compatible compressed-tensors build)

Responsibility

This model has had safety refusals removed and will comply with harmful requests. It is released for research into model alignment, red-teaming, and uncensored assistants. You are solely responsible for what you generate and how you use it; comply with all applicable laws. No warranty.

Provenance & credits

• Base: deepreinforce-ai/Ornith-1.0-35B (MIT)
• Driver: abliterix (Wangzhang Wu) · upstream heretic (Philipp Emanuel Weidmann)
• Methods: grimjim (norm-preserving biprojected abliteration), Arditi et al. 2024 (refusal direction), FernflowerAI (SSM conv1d repair)
• Build: AEON-7

License: MIT (inherited from the base model).