MiniMax-H3 FL2VA โ€” MLX-Serve, 2-bit text encoder / 4-bit DiT

Variant of the 4-bit MLX-Serve conversion of MiniMax-H3's FL2VA checkpoint: only the Qwen3-VL text encoder is quantized to MLX affine 2-bit (group size 64); the DiT stays affine 4-bit and the VAEs, tokenizer and config are byte-for-byte identical to the 4-bit release. Runs natively on Apple Silicon via mlx-serve. Text-to-audio-video: the DiT denoises video and stereo audio jointly in one packed sequence.

Component File Quantization
Qwen3-VL text encoder text_encoder.safetensors affine 2-bit, g64 (file metadata: affine 2-bit g64)
DiT (MiniMax-H3) transformer.safetensors affine 4-bit, g64
Video VAE video_vae.safetensors dense f16, unmodified
Audio VAE audio_vae.safetensors dense f32, unmodified
Tokenizer tokenizer.json etc. unmodified

How it loads: mlx-serve's MfLinear (shared by the H3 and MageFlow backends) derives (bits, group_size) per tensor from the packed geometry โ€” bits = 32ยทw_cols/in_features, group_size = in_features/scales_cols โ€” so each file is self-describing and no config change is needed. Every quantized tensor in text_encoder.safetensors solves to (2, 64); the DiT solves to (4, 64). The quantization block in config.json is advisory for this backend (reporting only) and still describes the DiT.

Same staged residency as the 4-bit release: the text encoder is loaded, run and freed before the DiT loads. The 2-bit encoder packs ~9.6 GB on disk vs ~15.8 GB at 4-bit.

Quality note: 2-bit affine is lossier than 4-bit by design. The text conditioning (prompt adherence) is where you may notice it; the DiT, VAEs and tokenizer are untouched, so frame/audio quality is otherwise the 4-bit release's.

Usage

Serve it with mlx-serve (a 32 GB Mac needs --skip-mem-preflight):

/Applications/MLX Core.app/Contents/MacOS/mlx-serve \
  --model /path/to/MiniMax-H3-FL2VA-MLX-Serve-2bit-text-encoder \
  --serve --host 127.0.0.1 --port 11235 --skip-mem-preflight

Then POST /v1/video/generations with {"prompt", "width", "height", "num_frames", "steps", "seed"} โ€” the response is base64 rgb8 frames plus pcm_s16le audio on the standard native wire format. A ready-to-run wrapper script is provided in the source directory of this conversion (h3gen.sh), with --te=2bit|4bit to switch between this variant and the original 4-bit release.

Modifications

These are MODIFIED files. The text encoder is quantized to 2-bit and the DiT to 4-bit; see MODIFICATIONS.md for the full list. The VAEs and tokenizer are unmodified copies.

License

Powered by MiniMax H3. Licensed under the MiniMax H3 Community License Agreement -- see LICENSE and NOTICE, both included here.

Territorial restriction. The Agreement defines the Applicable Territory as worldwide EXCLUDING the European Union, the United Kingdom, the Republic of Korea and the United States of America, and Section V.4 prohibits use, reproduction, modification, distribution and display outside it. Check whether your jurisdiction permits you to use these files before downloading them.

Downloads last month
-
MLX
Hardware compatibility
Log In to add your hardware

Quantized

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for antocorr/MiniMax-H3-FL2VA-MLX-Serve-2bit-text-encoder

Quantized
(32)
this model