--- license: mit base_model: deepreinforce-ai/Ornith-1.0-35B library_name: mlx pipeline_tag: text-generation tags: - mlx - qwen3.5-moe - vision-language - tool-use - dotresize - 8-bit --- # Ornith DOTResize 448 MLX oQ8 This is the oQ8 MLX quantized build, using affine group-size-64 quantization. The measured local artifact size is about 31 GiB. It is derived from `deepreinforce-ai/Ornith-1.0-35B` at revision `5df2ed3f675c7beaa490328cc70bb573b65fb660`. The model keeps the Qwen3.5-MoE/VLM structure and tokenizer/tool template, but reduces the MoE intermediate width from 512 to 448 with a support-preserving DOTResize pass. The goal of this build is to keep the 262k context metadata, coding behavior, and OpenAI-compatible tool calling while shaving memory from the original 35B BF16 artifact. Thinking is disabled unless a runtime explicitly supplies `enable_thinking=true`, which prevents normal chat responses from exposing `` delimiters or other internal wrapper tokens. The tool template is still present for runtimes that parse structured tool calls. ## Variants - BF16: `jedisct1/Ornith-DOTResize-448-MLX-BF16` - oQ8: `jedisct1/Ornith-DOTResize-448-MLX-oQ8` - oQ4: `jedisct1/Ornith-DOTResize-448-MLX-oQ4` ## Local Validation Before upload, the BF16 artifact passed static MLX safetensor inspection, oMLX lazy load, LM Studio load at 262144 context, direct OpenAI-compatible tool-call smoke, Swival core and all-tools gates at 5/5 with zero failed tool calls, a coding smoke, and a long-context retrieval smoke at 241,934 prompt tokens. The oQ8 and oQ4 artifacts were produced with oMLX streaming oQ quantization using `text_only=false`, then checked for VLM metadata preservation, quantized safetensor structure, oMLX lazy load, LM Studio load, direct tool-call behavior, and absence of visible thinking/tool wrapper tokens. ## Notes This is a derived local conversion and reduction, not a new base model. The validation above is practical smoke and integration coverage, not a replacement for a benchmark suite.