license: mit
base_model: AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4
library_name: transformers
pipeline_tag: text-generation
tags:
- uncensored
- abliterated
- qwen3_5_moe
- nvfp4
- moe
- reasoning
- vision
- dspark
- speculative-decoding
ORNITH-1.0_35B_AEON_PABLOG-OPTIMIZED_UNCENSORED_NVFP4
Base model (NVFP4) — the uncensored AEON Ornith-1.0-35B, paired with a DSpark speculative-decoding draft.
[!IMPORTANT]
🔗 TWO-PART MODEL — this is the BASE
This repo is the base model. For DSpark speculative decoding, you also need the matching draft (a separate download):
→ ORNITH-1.0_35B_AEON_PABLOG-OPTIMIZED_UNCENSORED_DSPARK-DRAFT_NVFP4
The base runs fine on its own for normal inference. The draft is optional and only accelerates this base (losslessly). To run the DSpark pair you download both.
What this is
The NVFP4 build of the uncensored / abliterated Ornith-1.0-35B (~22 GB), provided here paired with our DSpark draft. It is a re-host of the abliteration by AEON-7 of DeepReinforce's Ornith-1.0-35B — full credit to them for the base and abliteration. MIT-licensed.
- Architecture:
qwen3_5_moe— 40-layer GatedDeltaNet hybrid (30 linear-attention + 10 full-attention), 256 routed experts + 1 shared (A3B ≈ 3B active), vision tower, 256K context, reasoning (opens<think>). - Uncensored: refusal behavior removed via AEON's abliteration — 0/80 refusals on harmful-prompt probes, near-lossless (first-token KL ≈ 0.0014, identical agentic-coding pass@1).
- Precision: NVFP4-quantized (AEON's quantization) — ~22 GB, runs on vLLM 0.24.
- Measured: ~170 tok/s single-stream (base only, no draft) on 1× RTX PRO 6000 Blackwell.
⚠️ Uncensored disclaimer
This is an uncensored / abliterated model with refusal behavior removed; it will comply with requests a safety-tuned model would refuse. You are responsible for all generated content and for compliance with applicable law. For research and legitimate development use. Safety characteristics inherited from the AEON abliteration — see the AEON card.
Why we run the uncensored variant
- Post-hoc guardrails degrade quality — refusal training imposes an "alignment tax" (over-refusal, capability regressions bleeding into legitimate work); AEON's abliteration removes it while preserving capability (KL ≈ 0.0014).
- Guardrails belong to the deploying organization — controls depend on audience/domain/jurisdiction; an uncensored base is a neutral substrate for org-specific policy, not a vendor's fixed one.
- It performed better for us — in internal testing the AEON build outperformed the original Ornith-1.0-35B.
Deploy (vLLM)
vllm serve pablogrant/ORNITH-1.0_35B_AEON_PABLOG-OPTIMIZED_UNCENSORED_NVFP4 \
--served-model-name ornith --max-model-len 262144 \
--gpu-memory-utilization 0.9 --max-num-seqs 512 \
--mamba-cache-dtype float32 --reasoning-parser qwen3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coder \
--trust-remote-code
Reasoning model — opens <think>. Recommended sampling: temperature 0.6, top_p 0.95, top_k 20. Lower --max-num-seqs if you hit a Mamba-cache-block error (GatedDeltaNet hybrid).
Credits & license
- Abliteration: AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-NVFP4
- Upstream base: deepreinforce-ai/Ornith-1.0-35B
- DSpark draft (pair): → DSPARK-DRAFT_NVFP4
- Released under MIT.