Onyx · Subject Video MLLM encoder (Qwen3-VL-2B-Thinking, bf16)
The multimodal feature encoder for Onyx's Subject Video (S2V) pipeline on OnyxMac: caption +
reference images → [1, 1024, 2048] hidden-state features that guide the S2V DiT alongside the
umT5 text conditioning. Unmodified Qwen3-VL-2B-Thinking weights in the HF layout; Onyx's own
MLX extractor (S2VFeatureExtractor) loads this directory directly.
model.safetensors— 4.26 GB bf16 (model.visual.*+model.language_model.*).sha256.json— per-file integrity manifest.
License: Apache-2.0 (inherited from the base model).