MrMofer/MiniMax-H3-MLX-8bit

MiniMax-H3 MLX — 8-bit full pipeline (Apple Silicon)

MLX (Apple Silicon) build of the MiniMax-H3 33B joint video+audio diffusion transformer, packaged as a self-contained pipeline (DiT + text encoder + video VAE + audio VAE + tokenizer + configs) for the PipeNetwork MLX port.

Powered by MiniMax H3.

These files are modified. Quantized to 8-bit, text encoder rebuilt from the published MiniMax-H3-FL2VA-MLX-Serve-8bit checkpoint, tested and documented on an M4 Max. See PATCHES.md and the Validation report below.

Runtime pin

This repository was built and validated against:

• Base: PipeNetwork/minimax-h3-mlx commit b2f7e4d2 (2026-08-10)
• Patches (branch video-lab-8bit-fixes): 7210b93e6df86bf9c7206091c9542b6983c10c30 — see PATCHES.md
• MLX 0.32.0 (Python 3.12), mlx-vlm 0.6.13

Quick start

git clone https://github.com/PipeNetwork/minimax-h3-mlx
cd minimax-h3-mlx
git checkout b2f7e4d2   # base commit this repo was validated with
python3 -m venv .venv && .venv/bin/pip install -r requirements.txt

# Text-to-video+audio, validated settings (768x448, 16 sigma points -> 15 forwards)
.venv/bin/python scripts/generate.py "a red fox leaps over a mossy log in a misty forest" \
  -c <this repo> -t <this repo>/transformer -s 16 --seed 42 \
  --height 448 --width 768 -o fox.mp4

# Turbo LoRA (4-step EMA, included as turbo_lora_4step_ema.safetensors):
# merge it into a copy of the transformer with scripts/merge_h3_turbo.py,
# then run with -s 8 (see the Turbo LoRA section).

The pipeline needs only this repository: -c supplies the VAEs, text encoder and tokenizer; -t points at the quantized DiT. Native canvas is a 768px short edge (e.g. 1344x768 16:9); smaller canvases are off-distribution and degrade quickly.

Image-to-Video (FL2VA)

This repository is I2V-capable (partition fl2va, tasks t2va + fl2va in model_index.json). The image is optional — present as a keyframe conditioning row via --image / --anchor, or omitted for pure T2V.

Processor + vision tower required for I2V. The processor/ directory (7 files) and the 3 vision shards (text_encoder/model-00005..00007-of-00008-vision.safetensors + vision_quant_config.json, 529 model.visual.* tensors, 89 quantized) must be present. Without them the pipeline cannot encode images (falls back to T2V only).

Canonical I2V command (validated):

# 704×544 portrait (height 704 width 544) or 512×384 lightweight variant — both portrait 9:16 / 4:3
# Bridge image: inputs/bridge-20260813-181546-krea2_turbo-selfie-style-vertical-shot-a-young-woman-in-h.png
PROMPT=$(cat <<'PROMPT_EOF'
Vertical 9:16 TikTok-style UGC selfie video, handheld smartphone feel, natural indoor daylight near a window. A friendly creator speaks directly to camera with natural blinking, subtle head nods, and a warm smile. Add small human imperfections: a tiny hesitation, a soft breath, a quick smile mid-sentence, and a micro-pause before the last line. Realistic skin texture, stable identity, no face warping, minimal flicker, clean audio with natural room tone.

No subtitles. No on-screen text. No logos. No watermarks.

The creator says (exactly, with the same pacing and hesitations):
"Okay, entonces… eh… un datazo. Si estás trabado con tu código, solo da el primer pasito… como, abre tu terminal y escribe fran. (sonríe) Así de simple. Te vas a sorprender de lo rápido que todo se vuelve más fácil."
PROMPT_EOF
)

.venv/bin/python scripts/generate.py "$PROMPT" \
  -c <this repo> -t <this repo>/transformer -s 8 --seed 1996783985 \
  --height 704 --width 544 -d 5 \
  --image inputs/bridge-20260813-181546-krea2_turbo-selfie-style-vertical-shot-a-young-woman-in-h.png \
  --anchor first --release-encoder \
  -o i2v-704x544.mp4

# Lightweight alternative (faster, same identity):
.venv/bin/python scripts/generate.py "$PROMPT" \
  -c <this repo> -t <this repo>/transformer -s 8 --seed 1996783985 \
  --height 512 --width 384 -d 5 \
  --image inputs/bridge-20260813-181546-krea2_turbo-selfie-style-vertical-shot-a-young-woman-in-h.png \
  --anchor first --release-encoder \
  -o i2v-384x512.mp4

Valid anchors are first and last (FL2VA). Omit --image/--anchor for text-to-video. The image is consumed by text_encoder via processor/ (Qwen2VLImageProcessorPil fallback when torch absent) and the vision tower, then patched as video-conditioning rows (see minimal I2V encode in PATCHES.md).

Example outputs

Red fox on a mossy log in a misty forest (prompt above, seed 42). Columns are time (0.4s → 3.0s); rows: 8-bit 16-step, 8-bit + Turbo LoRA, 4-bit + Turbo LoRA at 1344x768.

Example clip (played inline by the Hub):

File	Build	Settings
examples/v3_768x448_16steps_8bit.mp4	8-bit (this repo)	768x448, 16 steps, 33.9 min

Turbo variants: see rows 2-3 above (a 1344x768 4-bit video re-render is planned).

Model layout

Path	Role	Representation
transformer/transformer.safetensors	33B DiT (joint video+audio)	8-bit affine, group 64 (AdaLN 8-bit), ~35.3 GB
text_encoder/ (5+3 shards)	Qwen3-VL-32B conditioner (64 layers, truncated to layer 50)	8-bit quantized, ~27.5 GB language (5 shards) + ~0.76 GB vision (3 shards model-00005..00007-of-00008-vision.safetensors, 529 model.visual.* tensors)
text_encoder/vision_quant_config.json	Vision tower quantization recipe (affine 8-bit g64)	json, 89 quantized tensors (blocks.*, merger.*, deepstack_merger_list.*)
processor/	Qwen3-VL processor (image preprocessing for I2V)	7 files: preprocessor_config.json, video_preprocessor_config.json, chat_template.json, tokenizer.json, tokenizer_config.json, vocab.json, merges.txt (~11.6 MB)
video_vae/	Tiled causal video VAE (17-frame chunks, latents_mean/std)	fp16/bf16, 5.2 GB
audio_vae/	DAC encoder + BigVGAN vocoder, stereo 32 kHz	fp32, 0.6 GB
tokenizer/	Qwen3-VL tokenizer	json/vocab/merges
model_index.json	Pipeline config (fl2va partition, sigma shift video 12.0 / audio 3.0)	json
turbo_lora_4step_ema.safetensors	4-step EMA Turbo LoRA (larryvrh, Apache-2.0)	758 MB
scripts/merge_h3_turbo.py	Merge the Turbo LoRA into a transformer copy	python

Memory and speed (measured, M4 Max 68.7 GB)

• Resident during generation: DiT ~21.5 GB + text encoder ~22 GB + VAEs ~6 GB (AdaLN projections are precomputed and dropped, freeing ~13.8 GB). With --release-encoder, the text encoder is freed after conditioning (prompt_embeds already built, before the denoise loop), freeing ~27.5 GB (log: released text encoder after conditioning (freeing ~27.5 GB) + mx.clear_cache() + tree_flatten sizing). Peak RSS drops by ~22 GB resident; overhead <1 s.
• 768x448, 16 steps: 33.9 min (119 s/step), peak RSS 24.7 GB, no swap. With turbo merge: 24 min (8 steps). With turbo8 (pure 8-bit, no merge) at 768x448, 8 steps (7 forwards): **18.5 min (139 s/step)** vs turbo merge 251 s/step — ~1.6–1.8× per-pixel speedup (same canvas, fewer denoise forwards).
• 1344x768 (native 16:9), 8 steps: ~2 h (996 s/step), peak ~57 GB + compressed memory; requires an otherwise idle machine (jetsam kills it under heavy ambient load).
• 704×544 I2V (portrait 9:16) turbo merge, 8 steps (7 forwards): ~33 min (251 s/step); turbo8: ~139 s/step (measured 768x448 baseline, scales linearly per pixel). Lightweight 384×512 I2V is proportionally faster (off-distribution, useful for wiring checks).
• TeaCache: measured 0 skips / 7 forwards with default last-block hook (--teacache, thresh 0.2, start 3, compute_last_step) at 768x448 turbo8 — probe feature pre-final-norm of last block (block 49/50) barely moves, so gate rel_l1 <= 0.2 never fires; with thresh 0.35 only 1/7 skips and still no wall saving (partial forward already traverses 49/50 blocks, only saves final norm + heads). Overhead +8–10% without skip. Recommendation: keep TeaCache OFF by default; if wall saving is needed, use an earlier hook --teacache-layer 40 with --teacache-thresh 0.25–0.35 (trades quality for ~20% block saving per skip). See PATCHES.md and tmp/S9_teacache_spike_RESULT.md.
• Attention is dense (MiniMax has not released sparse attention); quantization does not reduce the attention cost — it exists to fit memory.

Validation report (2026-08-16, M4 Max 68.7 GB)

Prompt used everywhere: "a red fox leaps over a mossy log in a misty forest", seed 42. Every clip was judged by an independent vision-capable model on extracted frames (not statistics alone):

Build	Settings	Time	Visual verdict
8-bit (this repo)	768x448, 16 steps	33.9 min	9/10 — coherent fox walking a mossy log; stable background; no melting
8-bit + Turbo LoRA	768x448, 8 steps	~24 min	High confidence — "impressive for a 4-step turbo LoRA"
4-bit + Turbo LoRA	1344x768, 8 steps	2 h 3 min	95% — full leap arc; no melting, no banding (frames; mp4 mux was a local script bug, not the model)
I2V E2E canonical (FL2VA, turbo merge)	544×704 (704h×544w) 9:16, 8 steps (7 forwards), 5 s, seed 1996783985, --image bridge-...png --anchor first --release-encoder	~33 min (251 s/step)	Keyframe anchored, identity preserved; requires processor/ + vision tower (529 tensors); image optional (FL2VA otherwise uses key)
I2V Turbo8 (pure 8-bit, FL2VA)	768×448, 8 steps (7 forwards), 5 s	~18.5 min (139 s/step, ~1.6–1.8× per-pixel vs turbo merge)	Same I2V prompt/image, pure 8-bit DiT without LoRA merge; faster, same wiring

Key findings:

• The VAE decoder is not the source of artifacts: parity with the reference diffusers implementation is exact (video 1.2e-6, audio 2.3e-8). The 17-frame chunked decode with 5-frame crossfade is the reference design; a periodic ~1.2x seam rhythm is architectural and masked at native resolution.
• Canvas is the dominant quality factor: the same VAE roundtrip scores 21.4 dB at 384x256 vs 27.9 dB at 768x448. Keep the short edge at 768.
• First-generation failures (flat field, collapsed latents) were traced to an incompatible third-party serve-pack conversion (inert condition_proj, exploding text-encoder activations), not to the pipeline; see PATCHES.md and the report in BIG_GOAL.md of the source workspace.

Turbo LoRA

turbo_lora_4step_ema.safetensors (also present as turbo_lora.safetensors — hardlink alias, same 779,849,816 bytes) is the 4-step EMA distilled LoRA from larryvrh/MiniMax-H3-Turbo-Lora (Apache-2.0; sha256 5a6eeba1…). Merge it into a transformer copy:

.venv/bin/python scripts/merge_h3_turbo.py --base transformer --lora turbo_lora_4step_ema.safetensors -o transformer-turbo
# then generate with the merged dir as -t and -s 8 (7 forwards)

8 steps denoise cleanly with the merged weights; 4 steps can collapse to noise (see port issue #5).

Patches vs upstream

PATCHES.md documents five patch families on branch video-lab over base commit b2f7e4d2 (2026-08-10): (a) DiT QKV blocked layout, (b) text-encoder 8-bit + processor fallback + positional scatter fix, (c) release_text_encoder headroom, (d) two-phase TeaCache, (e) I2V processor + vision shards. See PATCHES.md for per-file details and line references. The remaining pipeline is otherwise byte-identical to the base commit.

License

• Weights: MiniMax H3 Community License (included as LICENSE) — not open source; mark modified files, keep the attribution, territorial exclusions apply (EU/UK/KR/US excluded per the mlx-serve pack note). By downloading you accept those terms.
• Port code: Apache-2.0 (PipeNetwork/minimax-h3-mlx).
• Turbo LoRA: Apache-2.0 (larryvrh/MiniMax-H3-Turbo-Lora).