MiniMax-H3-FL2VA-MLX-BF16

MiniMax-H3 FL2VA MLX compact BF16

This is Mere's compact maximum-fidelity MLX runtime artifact for MiniMax-H3 FL2VA. Every weight input came from the pinned official checkpoint MiniMaxAI/MiniMax-H3@ec19cc6daf5d8add9417c18e86b6b58cc6c55027; no third-party converted or quantized checkpoint was used.

The 20.11B-parameter active denoising core remains BF16. The Qwen3-VL conditioner uses MLX affine INT8/group-64, the video VAE uses FP16, and the audio VAE remains FP32. Schedule-only AdaLN projections, the timestep MLP, and reconstructed RoPE tensors are omitted after generating a source-bound pack of exact 5, 9, 12, 16, 21, and 31-point tables at video/audio shifts 12/3 plus the LightX2V 5- and 9-point 6/3 tables.

The original cache pack was evaluated by mere.run model optimize on MLX Metal. The 9-point 6/3 table was added by the source-range refresh recorded in adaln_cache.refresh.json, also on MLX Metal. The refresh reproduced the tensor closure of every previously published table exactly before the new table was accepted. The pack remains bound to the official projection-tensor closure plus the original bit-identical 9- and 21-point real-generation receipts. The denoising core was reproduced on MLX CUDA; CUDA-evaluated AdaLN tables are not accepted because backend BF16 reduction order is not bit-identical to Apple Silicon.

SOURCE_MANIFEST.json, transformer.conversion.json, MODIFICATIONS.md, and SHA256SUMS provide the conversion and byte provenance. The artifact is intended for video-minimax-h3-fl2va-bf16-mlx in mere.run.

License

These weights are governed by the MiniMax H3 Community License Agreement in LICENSE, not the mere.run source license. At the pinned source revision the license excludes use, distribution, and display in the United States, European Union, United Kingdom, and Republic of Korea and carries additional notice and safeguard obligations. Review the complete license before downloading or using the artifact.