MiniMax H3 MLX — BF16 pruned transformer
Independent community conversion of the FL2VA pruned transformer. Two tested layouts: resident/ and stream2/ (two blocks per chunk). About 40.23 GB per layout; downloading both doubles storage. This repository contains transformer weights and the timestep curve; text encoder, video/audio VAEs and Turbo adapters are separate dependencies. It is not a generic Diffusers or MLX-LM model.
Runtime
git clone https://github.com/yshenaw/ComfyUI-MiniMax-H3-MLX-SolAttn.git
git -C ComfyUI-MiniMax-H3-MLX-SolAttn checkout 55bb7ddd015321a25385683419951e7c7f518775
git -C ComfyUI-MiniMax-H3-MLX-SolAttn apply /path/to/this-repo/runtime/local-fixes.patch
Place the runtime under ComfyUI custom_nodes, install its pinned requirements in the ComfyUI environment, and follow its setup instructions for the remaining dependencies. Tested with MLX 0.32.3 / PyTorch 2.14.1. Copy or download the contents of:
resident/→ComfyUI/models/minimax_h3/transformers/bf16-pruned/stream2/→ComfyUI/models/minimax_h3/transformers/bf16-pruned-stream2/
The tested ComfyUI node checks resident metadata/weights during preflight even when streaming; download both layouts for that exact integration. A standalone StrictStagedTextToVideo pipeline can use stream2/ directly. Select memory mode stream2, I/O offset, and Qwen stages 2 in the supplied workflow.
Shared dependencies: official MiniMax H3 FL2VA VAEs, Qwen3-VL-32B MLX 8-bit encoder, and Larry Turbo adapters. Use v1 ckpt850 for four steps; v4 step600 EMA for the included mixed six-step alternative.
What changed
Quantization/streaming use upstream tooling with local fixes. The patch preserves precision overrides in streamed blocks, fixes Torch→MLX storage ownership during conversion, exposes two-stage text execution and the v4 adapter, and releases inference caches after export. No fine-tuning was performed. Source, revisions and curve provenance are in PROVENANCE.json; precise precision settings are in resident/quant_config.json when quantized.
Local qualification
October 7, 2026: M4 Max / 64 GiB, ComfyUI 0.39.0, 864×480, 124 frames / 24 fps, synchronized H.264/AAC. 1885.9 s server execution for the supplied configuration. This is a historical single-scene measurement under shared host activity, not an averaged benchmark or a new public-revision inference test. validation.json records limitations; api_workflow.json records parameters. Core4 and mixed tests retained the LLM; the mixed v4 alternative overlapped lifecycle tests and does not establish uninterrupted LLM coexistence.
The local patch had 41 focused schema, recipe, streaming and conversion checks pass. Generated media were inspected for coherent cup appearance and finite synchronized audio. Runtime is text-to-video with audio; this integration does not expose image/keyframe conditioning, despite broader parent capabilities. Results trade speed and quality through precision, Turbo, sparse attention and cache choices.
Sample is AI-generated with MiniMax H3. Verify downloads with SHA256SUMS. RELEASE_MANIFEST.json records exact published files.
License and attribution
Model files retain the MiniMax H3 Community License; see LICENSE and NOTICE. Download acceptance binds recipients to LICENSE/AUP. Follow MiniMax’s licensing process where territorial authorization is required. Official distribution clarification. This release grants no additional model rights. Runtime code retains Apache-2.0 (CODE_LICENSE).
Credits: MiniMax, Comfy-Org, Larryvrh, yshenaw’s MLX/Sol-Attn runtime, lmstudio-community. Conversion, local fixes and packaging: Shiftedx. No publisher endorsement.