v0.0.1a2

Second alpha pre-release of the pure MLX MiniMax-H3 inference runtime for Apple Silicon.

Highlights

• Add first-frame and first/last-frame conditioning through the FL2VA path.
• Add ordered Ref2VA image, video, standalone-audio, and video-soundtrack references.
• Add the pure MLX Qwen3-VL vision tower plus Video VAE and Audio VAE encoders.
• Add schedule-aware exact AdaLN precompute, reducing active DiT residency from about 34.8 GiB to 21.2 GiB before denoising.
• Add a minimal public Python API and --prompt-file support for long UTF-8 requests.
• Validate all required tokenizer and checkpoint paths before loading a model.

Validation

• 84 tests pass; 34 optional checkpoint/runtime tests skip when their local assets are absent.
• Ruff, public-tree, whitespace, wheel, sdist, and clean-environment installation checks pass.
• Five-second 10-step reference previews remain near a 28.0 GiB global peak with no swap growth on the development machine.

Model weights and tokenizer assets are not included. Inference uses the 8-bit DiT and text encoder; dense DiT/text checkpoints are offline requantization inputs only.

v0.0.1a1

First public pre-release of the pure MLX MiniMax-H3 inference runtime for Apple Silicon.