language:
- en
license: apache-2.0
base_model: ornith-ai/Ornith-1.5-35B-A3B
library_name: mlx
pipeline_tag: image-text-to-text
tags: - mlx
- apple-silicon
- uncensored
- abliterated
- crack
- jang
- mxfp8
- reasoning
- vision
- video
- agentic-coding
- moe
- harmbench
- mmlu
- ornith
Built for vMLX — the MLX inference engine for Apple Silicon with mixed-precision JANG bundles, KV-cache quantization, and agentic tool calling.
Free for macOS · vmlx.net
Ornith 1.5 35B — UNCENSORED CRACK
MXFP8 · 8-bit MXFP8 (near-lossless reference)
Uncensored · Vision + Video · Reasoning on by default · Agentic coding · 262K context · ~35 GB
What Is This?
ornith-ai/Ornith-1.5-35B-A3B — a 35.9B
Mixture-of-Experts vision-language model (40 layers, 256 routed experts, hybrid gated-delta +
full-attention backbone, 27-layer vision tower, native video) — uncensored and quantized to a
8-bit MXFP8 (near-lossless reference) MLX bundle for Apple Silicon.
Refusal behavior is removed at the weight level: the model follows instructions across task
categories instead of refusing, while keeping its coding ability, knowledge, reasoning, and vision
intact. No runtime hooks, no steering vectors — a standard MLX bundle.
Results
Measured on this exact bundle. MMLU is the standard 57-subject benchmark in logit mode. HarmBench
compliance is coherence-gated (looping or template dumps do not count) and excludes
copyright-reproduction behaviors. KL divergence is measured against the uncracked MXFP8
reference on neutral held-out text — lower means closer to the original model's behavior.
| Metric | Value |
|---|---|
| MMLU (57-subject) | 78.9% (base 80.6%, -1.67) |
| HarmBench compliance | 100.0% (240/240) |
| KL vs uncracked MXFP8 | 0.0289 nats (floor 0.0000) |
| Size | ~35 GB |
MMLU by category — base vs uncensored
| Category | Base | Uncensored | Δ |
|---|---|---|---|
| STEM | 75.8% | 73.2% | -2.6 |
| Humanities | 81.5% | 81.2% | -0.4 |
| Social Sciences | 87.5% | 87.9% | +0.4 |
| Other | 80.4% | 76.9% | -3.5 |
| Overall (57 subj) | 80.6% | 78.9% | -1.67 |
Capability is preserved: the model stays within a few points of the base bundle at the same
quantization while refusals are removed.
Modalities
| Vision | supported — pass images through the bundled processor |
| Video | supported (native video preprocessor) |
| Reasoning | on by default; toggle with enable_thinking |
| Tool calling | native XML / function schema |
| Context | 262,144 |
Usage
Run with vMLX (recommended — honors the per-module mixed-precision overrides)
or an MLX-VLM runtime with qwen3_5_moe support.
Recommended sampling (coding preset): temperature 0.6, top_p 0.95, top_k 20. A general preset
(temperature 1.0) is also stamped in jang_config.json. Stop tokenseos_token_id = [248046, 248044].
{
"model": "dealignai/Ornith-1.5-35B-A3B-MXFP8-UNCENSORED-CRACK",
"messages": [{"role": "user", "content": "..."}],
"temperature": 0.6, "top_p": 0.95, "top_k": 20,
"enable_thinking": true
}
Support dealignai
Support us on Ko-fi · X @dealignai · dealign.ai
⚠️ Disclaimer
This model has had its safety-refusal behavior removed for research purposes. It will follow
instructions across all categories without refusing. You are solely responsible for how you use it
and for complying with all applicable laws. Published for AI-safety research and authorized
security testing.