license: apache-2.0
base_model: AEON-7/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16
tags:
- openvino
- int4
- intel-arc
- qwen3_5
- mixture-of-experts
- roleplay
- uncensored
language: - en
Ornith-1.0-35B (Qwen3.6-35B-A3B), OpenVINO INT4. 106 tok/s on ONE Intel Arc B70.
This is AEON-7's Ornith "Ultimate Uncensored" tune of Qwen3.6-35B-A3B (256-expert MoE, 3B
active parameters), converted to OpenVINO INT4 with Intel's own recipe. Sibling of my
tvall43 heretic conversion
of the same base model: same speed class, different tuning flavor. Runs on the stock 2026.2
runtime, no patches, coherent through 32K context.
One thing made this conversion special. Ornith was saved by a bleeding-edge transformers
(5.13-dev) that stores each of the 10,240 experts as separate tensors (30,720 tensor files
worth of them), while every current export toolchain expects the fused layout and silently
RANDOMIZES the experts if it does not find it. This conversion required re-fusing all 30,720
per-expert tensors back into the 80 fused ones, with the layout and packing order verified
against the toolchain's own model skeleton before writing. If you are trying to convert any
model saved in the new per-expert format, the fusion script and the full story are in the
toolkit repo. But, you should know
I did all that work to find out it was pretty tarded for roleplay, still good for workflow,
but not the best at campfire stories.
Measured performance (single Arc Pro B70, OpenVINO 2026.2)
| Metric | Value |
|---|---|
| Decode, short context | ~106 tok/s |
| Decode at 6K context | ~77 tok/s |
| Prefill | pp512 ~1,100 tok/s; ~3,900 at 6K; 16K in ~7 s; 32K in ~22 s (curve measured on the tvall43 twin, same architecture) |
| Model load | ~16 s |
| Weights | ~19 GB (needs a 24 GB+ card; B60/B70 class) |
Long-context capability, verified by needle retrieval
A password fact planted early in the prompt, retrieved at the end. Pass means the exact password.
| Context | Result |
|---|---|
| 8K | PASS (2.5 s total) |
| 16K | PASS (7.7 s) |
| 24K | PASS (19.0 s) |
| 32K | PASS (41.7 s, cold cache) |
No rope patch, no precision workarounds, no special properties for single-stream use. The
architecture is rated to 262K positions; 32K is as far as my test box's host RAM lets me verify.
How to run
Identical to the tvall43 sibling:
standard Qwen <|im_start|> format, pre-closed <think>\n\n</think> block for fast no-think
replies (remove it and give at least 1024 tokens to enable reasoning), and{"DYNAMIC_QUANTIZATION_GROUP_SIZE": 0} if you use continuous batching. Full runnable snippet
on the sibling card; just point it at this folder.
pip install openvino-genai==2026.2.0 huggingface_hub
huggingface-cli download Wondernutts/Ornith-1.0-35B-AEON-Ultimate-Uncensored-BF16-int4-ov --local-dir ./ornith-35b-ov
Provenance
Qwen/Qwen3.6-35B-A3B, Ornith-1.0 "AEON Ultimate Uncensored" tune by
AEON-7,
per-expert-to-fused weight restoration and OpenVINO INT4 conversion (this repo).
Intended use and content notice
Uncensored general model, built and tested for roleplay and creative writing on local Intel
hardware. The tune removes refusal behavior and outputs are unfiltered; you are responsible for
lawful and appropriate use. Licensed under
Apache 2.0, same as the upstream Qwen3.6 release.