Ornith 1.0 35B abliterated — vLLM-ready mirror
Hosted inference: run this model on demand at imbutus.com.
A drop-in, serving-ready mirror of
YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated.
The weights are identical to the source repo — the only change is that the
missing multimodal preprocessor configuration files were added so the model
loads on vLLM (and any transformers-based multimodal loader) without patching.
Why this mirror exists
config.json declares a multimodal model:
"architectures": ["Qwen3_5MoeForConditionalGeneration"],
"model_type": "qwen3_5_moe",
"vision_config": { ... },
"image_token_id": ..., "video_token_id": ...
so vLLM initializes the vision path and requires an image/video processor config. The source abliterated re-upload did not include those files, so loading it directly fails with:
OSError: Can't load image processor for
'YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated' ... make sure ...
contains a preprocessor_config.json file
What was added
Copied verbatim from the official base model
deepreinforce-ai/Ornith-1.0-35B:
preprocessor_config.jsonprocessor_config.jsonvideo_preprocessor_config.jsonvocab.json
Everything else (the two safetensors shards, config.json, tokenizer.json,
chat_template.jinja, etc.) is a byte-for-byte copy of the source repo.
Serving with vLLM
vllm serve imbutus/Ornith-1.0-35B-abliterated \
--trust-remote-code \
--max-model-len 131072 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice --tool-call-parser qwen3_coder
Fits a single 96 GB GPU (~70 GB bf16 weights). Native context 262144, capped here to 131072 for single-GPU KV cache.
Credits
- Base model:
deepreinforce-ai/Ornith-1.0-35B - Abliterated weights:
YuYu1015/YuYu1015-Ornith-1.0-35B-abliterated
This mirror only re-adds the standard preprocessor configs; no weights were modified.