{"slug": "antirez-h3-c-minimax-h3-inference-engine-for-mac-computers", "title": "Antirez/h3.c: MiniMax H3 inference engine for Mac computers", "summary": "Antirez released h3.c, a native MiniMax-H3 inference engine for Apple Silicon Macs, with prompt-to-video/audio, first/last-frame conditioning, and ordered Ref2VA references working end to end. The project currently focuses on incremental H3-specific Metal performance and memory optimization on M3 Max and M5 Max, and supports interactive sessions and a balanced preset generating 22 frames at 24 fps (about 0.92 seconds) with configurable steps, layers, and reuse.", "body_md": "Native MiniMax-H3 inference for Apple Silicon. The project is being built as a sequence of working vertical slices: deterministic host/model metadata first, then portable Metal block parity, prompt encoding, prompt-to-video/audio, and first/last-frame conditioning and then ordered references.\n\nPrompt-to-video/audio, first/last-frame conditioning, and ordered Ref2VA image/video/audio references work end to end. The current work is incremental H3-specific Metal performance and memory optimization on M3 Max and M5 Max.\n\nThe examples assume that the Hugging Face snapshot is in `./MiniMax-H3`\n\nand\nthat FFmpeg and FFprobe are available on `PATH`\n\n.\n\n```\nmake -j8\nmkdir -p outputs\n./h3 --info -d ./MiniMax-H3\n```\n\n`--info`\n\nchecks the model layout and prints the selected Metal device without\nmapping all weights or generating media. Run `./h3 --help`\n\nfor the complete CLI\nreference.\n\nWithout `-p`\n\n, the same binary starts an Iris-style interactive session:\n\n```\n./h3 -d ./MiniMax-H3 --width 512 --height 512 --steps 6\n```\n\nType a prompt to generate a numbered video. The session keeps the exact BF16\nprompt conditioning, prepared DiT, and video decoder in memory, so repeating a\nprompt with another seed avoids loading and encoding them again. Useful commands\nare `!status`\n\n, `!seed random`\n\n, `!seconds 2`\n\n, `!show`\n\n, `!save output.mp4`\n\n, and\n`!cache`\n\n. Use `!help`\n\nfor the full, short list.\n\nFirst/last-frame conditioning is persistent in the session:\n\n```\nh3> !first opening.png\nh3> !last ending.png\nh3> The camera moves slowly around the subject.\n```\n\nUse `!first clear`\n\nor `!last clear`\n\nto remove an anchor. Generated videos are\nwritten to the session directory printed at startup.\n\nFor a general Ref2VA conditioning image, use `!ref-image PATH`\n\ninstead. Images\nare appended in order and exposed to the model as ``\n\n, ``\n\n,\nand so on; filenames have no meaning to the model.\n\n```\nh3> !ref-image person.png\nh3> Make the person shown in Picture 1 wave to the camera.\n```\n\n`!refs`\n\nlists the current order, `!ref-remove N`\n\nremoves one entry, and\n`!refs clear`\n\nremoves them all. Ref2VA references cannot be mixed with\n`!first`\n\n/`!last`\n\nanchors.\n\nStart with the validated balanced preset. It generates 22 frames at 24 fps (about 0.92 seconds), displays the evolving middle-video frame after every denoising transition in a supported graphical terminal, and prints phase timings:\n\n```\n./h3 --profile \\\n -d ./MiniMax-H3 \\\n -p \"A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur, soft footsteps and wind.\" \\\n --width 512 --height 512 \\\n --frames 22 --steps 20 \\\n --layers 45 --reuse 2 \\\n --show \\\n -o outputs/fox-fast.mp4\n```\n\nThis is deliberately not the most aggressive configuration:\n\n`--steps 20`\n\nperforms the default 20 denoising passes.`--reuse 2`\n\ncomputes 11 fresh denoiser velocities instead of all 20 and extrapolates the skipped transitions.`--layers 45`\n\nruns 45 of the 50 transformer blocks, reducing both time and unified-memory use.`--show`\n\nis optional. It supports Kitty/Ghostty and iTerm2/WezTerm/Konsole graphical protocols. It loads a resident preview VAE, displays one representative middle-video frame after every Euler transition, and then displays all final frames. Display dimensions default to 2x so the image has its intended logical size on macOS Retina screens; use`--zoom 1`\n\non a non-HiDPI display. This adds preview decode time and roughly 10 GiB of temporary model residency; runs without`--show`\n\nare unchanged.`--profile`\n\nis optional and does not select a different generation path.\n\nThe first process invocation also pays model loading and filesystem-cache costs. Compare performance using repeated runs, and alternate variants when the machines are warming up because this workload is sensitive to thermal throttling.\n\nFor a very short iteration, request four denoising passes directly:\n\n```\n./h3 --profile \\\n -d ./MiniMax-H3 \\\n -p \"A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur.\" \\\n --width 512 --height 512 --frames 22 \\\n --steps 4 --layers 50 --reuse 1 \\\n --show \\\n -o outputs/fox-four-step.mp4\n```\n\n`--steps N`\n\nalways means exactly N denoising passes. Four through seven passes\nuse the same schedule that won the low-budget comparison; increasing from 4\nto 7 progressively improves detail and motion. Keep `--reuse 1`\n\nat such small\nbudgets so every requested pass runs the model. `--show`\n\ndisplays one preview\nafter each pass.\n\nSeveral tail-heavy schedules were evaluated because most visible cleanup happens late in a long run. They preserved too few early composition updates and produced woven texture, weak motion, or clipped colors. The retained mode uses the released linear base grid with one terminal point. On the 512-square, 22-frame fox test, the selected four-pass result had 0.556 full-video SSIM against a 29-pass reference; an independent surfer test measured 0.547. The four-pass denoise took about 3.5 seconds on M5 Max, versus 26.4 seconds for the reference.\n\nChange one control at a time when evaluating quality. First restore all layers, then all denoiser evaluations, and finally raise the default 20-pass schedule to the slower 50-pass reference:\n\n```\n./h3 --profile \\\n -d ./MiniMax-H3 \\\n -p \"A red fox walks through fresh snow in a pine forest. Medium tracking shot, natural winter light, realistic fur, soft footsteps and wind.\" \\\n --width 512 --height 512 \\\n --frames 22 --steps 50 \\\n --layers 50 --reuse 1 \\\n -o outputs/fox-close.mp4\n```\n\nThe defaults are `--steps 20 --layers 50 --reuse 1`\n\n; keep `--steps 50`\n\nexplicit for this close path. It performs 50 complete 50-block denoiser\nforwards and is much more expensive than the default, but is the right oracle\nwhen a fast mode changes the subject, anatomy, motion, or composition.\nNumerical pixel identity with MLX is not expected because the random-number and\nexecution engines differ; the depicted content and motion should agree.\n\nThese controls are independent unless noted otherwise:\n\n| Control | Slow reference | Default | Aggressive | Main impact |\n|---|---|---|---|---|\n| Denoising passes | `--steps 50` |\n`--steps 20` |\n`--steps 4..7` |\nThe number always names actual denoising passes. |\n| Whole denoiser reuse | `--reuse 1` |\n`--reuse 2` |\n`--reuse 3` |\nAt 20 steps: 20, 11, or 8 fresh DiT evaluations. |\n| Active DiT blocks | `--layers 50` |\n`--layers 45` |\n`--layers 40` |\nFewer blocks reduce compute and resident transformer weights. |\n| Core residual reuse | `--core-reuse 1` |\n`--core-reuse 4` |\n`--core-reuse 6` |\nRefreshes patch/head work every step but runs the expensive core less often. |\n| Token reduction | off | optional | `--token-reduction` |\nPairs horizontal video tokens inside middle blocks; faster but may change composition. |\n| Internal canvas | output size | `384x384` for 512 square output |\n`320x320` |\nRuns DiT/VAE smaller, then upscales with vImage. |\n\nOn M5, `--use-int8-row-fc2`\n\nuses one activation scale per FC2 row and a single\nfull-width TensorOps product. It is optional because it is less numerically\nconservative than grouped int8. It reduced complete denoiser forwards by about\n2.6% in reciprocal tests. Matched four-step fox and surfer videos kept the same\nsubjects, setting, and motion (full-video SSIM 0.919 and 0.828). In the\ninteractive session, use `!int8-row-fc2 on`\n\n.\n\n`--reuse`\n\nand `--core-reuse`\n\nare mutually exclusive. Layer thinning can be\ncombined with either one.\n\nTo make the first command faster while keeping its output resolution, add token reduction:\n\n```\n./h3 --profile \\\n -d ./MiniMax-H3 \\\n -p \"A surfer riding inside a sharp blue ocean wave, one rider and one white board, realistic spray.\" \\\n --width 512 --height 512 --frames 22 --steps 20 \\\n --layers 45 --reuse 2 --token-reduction \\\n -o outputs/surfer-fast.mp4\n```\n\nAt the validated 512 square shape, token reduction cut the `45 layers + reuse 2`\n\ndenoise profile from 16.69 to 12.60 seconds on the IT M5 Max. Independent\nfox and surfer renders stayed coherent, but composition can diverge more from\nthe close path.\n\nFor an aggressive preview, render internally at 320 square and upscale to the requested 512 square output:\n\n```\n./h3 --profile \\\n -d ./MiniMax-H3 \\\n -p \"A red fox walking through snow, realistic, tracking shot.\" \\\n --width 512 --height 512 \\\n --render-width 320 --render-height 320 \\\n --frames 22 --steps 20 --layers 40 --reuse 3 \\\n -o outputs/fox-aggressive.mp4\n```\n\nThis combination produced a clean, recognizable 22-frame fox in validation,\nbut loses fine detail and can change framing. Do **not** add `--token-reduction`\n\nto both `--layers 40`\n\nand `--reuse 3`\n\n: that tested combination produced color\nringing, outlines, and ghosted limbs.\n\nAs an alternative to whole-velocity reuse, this keeps the timestep-dependent patch and output heads fresh at every transition:\n\n```\n./h3 --profile \\\n -d ./MiniMax-H3 \\\n -p \"A surfer riding a blue ocean wave.\" \\\n --width 512 --height 512 --frames 22 --steps 20 \\\n --layers 45 --core-reuse 4 \\\n -o outputs/surfer-core-reuse.mp4\n```\n\nUse `--core-reuse 6`\n\nonly as an aggressive preview. Values above 6 are not\nexposed because validation lost subject fidelity.\n\nWidth and height must each be multiples of 32, at least 32, and their product\nmust not exceed `768 * 1344`\n\npixels. Those are mechanical limits, not a promise\nthat every tiny canvas has good model quality. H3-Base is a 768p model.\n\n| Canvas | Current guidance |\n|---|---|\n`512x512` |\nSafest development size; repeatedly validated with multiple prompts. |\n`768x768` |\nValidated close-quality square output; substantially more expensive. |\n`1344x768` , `768x1344` |\nReleased 768p-class landscape/portrait limit. |\n`1024x768` , `768x1024` |\nValid 4:3 and 3:4 768p-class canvases. |\n`384x384` internal to `512x512` |\nValidated fast-quality scaling point. |\n`320x320` internal to `512x512` |\nValidated aggressive scaling point. |\n`256x256` |\nNative fast-preview canvas with automatic low-resolution RoPE adaptation. |\n\nFor a fast native 256-square preview:\n\n```\n./h3 -d ./MiniMax-H3 \\\n -p \"A red fox walks through fresh snow in a pine forest.\" \\\n --width 256 --height 256 \\\n --frames 22 --steps 20 \\\n --layers 50 --reuse 1 \\\n -o outputs/fox-256.mp4\n```\n\nAt 256 square, H3 has only an `8x8`\n\neffective spatial-token grid, so it has less\nroom for fine detail and complex composition. H3 automatically halves spatial\nRoPE coordinates at exactly 256 square. This removed repeating lattice\nartifacts in long fox renders and stayed coherent on an independent portrait,\nwithout adding tokens or runtime. Use `--use-reference-rope`\n\nto restore the\nreleased/MLX coordinates for parity checks. Keep token reduction off at this\nsize. Native 128 square remains unsupported: its `4x4`\n\ntoken grid did not\nrecover a recognizable subject even with adjusted RoPE.\n\n`--render-width`\n\nand `--render-height`\n\nmust be set together, must have the same\naspect ratio as the output, and cannot exceed the output dimensions. The model\nand VAE use the internal size; terminal frames and the encoded video retain the\nrequested output size.\n\nH3 emits 24 fps and aligns frame requests upward to `5 + 17*n`\n\n:\n\nUse `--seconds N`\n\nfor a duration-oriented request, or `--frames N`\n\nfor direct\nframe control; the two options are mutually exclusive. Fractional seconds are\naccepted. Seconds are converted at 24 fps and then rounded upward to the next\nlegal H3 temporal shape, so `--seconds 10`\n\nproduces 243 frames (10.125 seconds).\n\n| Frames | Approximate video duration |\n|---|---|\n| 22 | 0.917 seconds |\n| 39 | 1.625 seconds |\n| 56 | 2.333 seconds |\n| 107 | 4.458 seconds |\n| 243 | 10.125 seconds |\n| 362 | 15.083 seconds |\n\nShort clips are useful for development. The released workflow is intended for\nroughly 4–15 second videos. A request such as `--frames 23`\n\nis rounded up to 39\nframes rather than producing an arbitrary temporal shape.\n\nA short prompt works, but the released system expects a Context-IR-like description. State the subject, action, setting, camera, lighting/style, and desired sound. For example:\n\n```\nScene: a single red fox in a snow-covered pine forest at dawn.\nAction: the fox walks steadily left to right and looks toward the camera once.\nCamera: medium-height lateral tracking shot, 50 mm lens, stable framing.\nLook: photorealistic fur, cold blue ambient light, warm sunrise rim light.\nAudio: soft footsteps in snow, light wind through pine branches, no music.\n```\n\nKeep identity and object counts explicit when they matter. `--seed N`\n\ncontrols\nthe native random stream; the default is 42. Compare options with the same\nprompt, seed, resolution, frame count, and step count.\n\n`--show`\n\ndisplays a representative frame after every denoising transition, followed by all frames from the completed video. Like Iris, it advertises 2x display dimensions by default for Retina terminals;`--zoom N`\n\nchanges that factor without resizing the generated video or the encoded terminal image.`--frames-dir DIR`\n\nwrites final callback frames as PPM files. Intermediate`--show`\n\npreviews are not written there.`-o ''`\n\ndisables MP4 encoding; combine it with`--frames-dir`\n\nwhen FFmpeg is unavailable.`--profile`\n\nreports phase wall time, Metal encoding/wait time, peak live tensor storage, cumulative allocation, and dispatch counts.\n\nFor example:\n\n```\n./h3 --profile -d ./MiniMax-H3 -p \"A hummingbird hovering over red flowers.\" \\\n --width 512 --height 512 --frames 22 --steps 20 \\\n --layers 45 --reuse 2 --frames-dir outputs/hummingbird-frames \\\n -o ''\n```\n\nFirst/last-frame anchors select the FL2VA path:\n\n```\n./h3 -d ./MiniMax-H3 -p \"The fox keeps walking through the snow.\" \\\n --width 512 --height 512 --frames 22 --steps 20 \\\n --layers 45 --reuse 2 \\\n --first-frame fox.png --last-frame fox-later.png \\\n -o outputs/fox-anchored.mp4\n```\n\nOrdered references select the distinct Ref2VA checkpoint. Use the flag matching the media semantics:\n\n```\n# One image reference.\n./h3 -d ./MiniMax-H3 -p \"Use the animal and setting in the reference.\" \\\n --width 512 --height 512 --frames 22 --steps 20 \\\n --ref-image fox.png -o outputs/fox-reference.mp4\n\n# Continue a clip but ignore its soundtrack.\n./h3 -d ./MiniMax-H3 -p \"Continue the motion in this clip.\" \\\n --width 512 --height 512 --frames 22 --steps 20 \\\n --ref-silent-video fox.mp4 -o outputs/fox-video-reference.mp4\n\n# Preserve the clip's embedded audio.\n./h3 -d ./MiniMax-H3 -p \"Continue this audiovisual scene.\" \\\n --width 512 --height 512 --frames 56 --steps 20 \\\n --ref-video fox-with-audio.mp4 -o outputs/fox-video-audio.mp4\n\n# Replace a video's soundtrack explicitly.\n./h3 -d ./MiniMax-H3 -p \"Continue the scene with the supplied music.\" \\\n --width 512 --height 512 --frames 56 --steps 20 \\\n --ref-video-audio silent-fox.mp4 replacement.wav \\\n -o outputs/fox-replaced-audio.mp4\n\n# An ordered image plus standalone audio reference.\n./h3 -d ./MiniMax-H3 -p \"Use the animal and music from the references.\" \\\n --width 512 --height 512 --frames 56 --steps 20 \\\n --ref-image fox.png --ref-audio music.wav \\\n -o outputs/fox-image-audio.mp4\n```\n\nReference flags may be repeated and their command-line order is preserved. Standalone audio must accompany an image or video reference. Audio references must be 2–15 seconds; at most three audio inputs are accepted and their total decoded duration is capped at 15 seconds.\n\n```\nmake test\nmake parity\n```\n\n`make test`\n\nruns the deterministic host suite and, when the ignored MLX fixture\nis installed under `misc/fixtures/`\n\n, compiles the Metal source at runtime and\nchecks a complete toy H3 block against named MLX outputs. Runtime compilation is\nintentional: it follows Iris and does not require Xcode's optional offline Metal\ntoolchain. The test covers both an F32 diagnosis path and the production BF16\nstorage path; wide BF16 matrix products and SDPA use cached MPSGraph graphs, with\ndirect Metal correctness fallbacks. `make parity`\n\nruns only those Metal/MLX\nchecks.\n\nFFmpeg and FFprobe must be available on `PATH`\n\nfor media inputs and MP4 output\n(`H3_FFMPEG`\n\nand `H3_FFPROBE`\n\nmay select explicit executables). Generated RGB24 and\n32 kHz stereo F32 PCM are fed through concurrent pipes; no intermediate\nuncompressed media file is created.\n\nThe remainder documents the implementation behind the tutorial presets and the environment variables retained for exact A/B diagnosis.\n\nThe default sampler uses the released shifted video/audio schedule. `--steps`\n\nalways names the number of denoising passes, with terminal zero added after the\nlast pass. Whole-denoiser reuse evaluates the first and last pass plus every\nrequested interval, then extrapolates skipped video and audio velocities on\ntheir independent schedules. With very small step counts, keep `--reuse 1`\n\n.\n\nFor the low-budget path, the released linear base grid won against actual-video-sigma linear spacing, quadratic and cubic warps, exact 30-point tail subsets, mild power warps, zero-order held full-grid velocities, linear velocity extrapolation, and RES. The more tail-heavy candidates often sharpened the subject but damaged motion or left a repetitive woven background; sparse RES and long extrapolation intervals failed much more visibly.\n\nLayer thinning ranks the checkpoint's actual AdaLN gates while protecting\nstructurally important first and final blocks. Unused weights and schedule\ntensors are not retained, so `--layers 45`\n\nand `--layers 40`\n\nreduce both\ntransformer time and unified-memory use. Core reuse holds the previous full\ntransformer residual while refreshing the patch projection and timestep-aware\nhead; it remains mutually exclusive with whole-velocity reuse.\n\nEvery active DiT block fuses its attention residual gate with the following MLP\nAdaLN. The rounded BF16 residual is still written exactly, but the same row is\nkept in threadgroup memory for normalization, eliminating one dispatch and one\nglobal reread. Away from token-reduction boundaries, the MLP residual gate also\nproduces the next block's attention AdaLN and carries that normalized state\nacross the loop. `H3_DISABLE_FUSED_GATE_ADALN=1`\n\nand\n`H3_DISABLE_FUSED_CROSS_BLOCK_ADALN=1`\n\nrestore the two-kernel oracles.\nThe final audio/video AdaLN kernels bind directly to offsets in the residual\nstream, avoiding two slice blits and 18.8 MiB of scratch at 512x512 (29.4 MiB\nat the 864-class benchmark shape).\n`H3_DISABLE_FUSED_FINAL_SLICE=1`\n\nrestores the copy-plus-AdaLN oracle at load.\nThe BF16 final heads then apply AdaLN while loading their 16x16 projection\ntiles, preserving the standalone rounding and accumulation order while\nremoving another equally sized normalized activation. The two optimizations\ntogether save 37.5/58.9 MiB. `H3_DISABLE_FUSED_FINAL_HEAD=1`\n\nrestores the\noffset-AdaLN-plus-linear oracle at load.\n\n`--token-reduction`\n\nis an independent aggressive DiT mode. After block 3 it\npairs adjacent horizontal target-video tokens while leaving text, audio,\nconditions, and reference tokens exact. The complete full-resolution state is\nkept as a bypass. During the first ten noisy evaluations it restores before\nblock 40; subsequent detail-forming evaluations restore before block 30. Each\ntoken returns as its original value plus the update learned by its pair, so\nwithin-pair detail is not discarded.\nThe pooling kernel writes only true-pair baselines into a dense tail of the\nalready allocated attention scratch buffer; odd-width singleton tokens need no\nbaseline. The full bypass uses the oversized QKV tail when it fits, with a\nguarded dedicated fallback only for reference-heavy layouts. Common text-only\ncanvases therefore add no activation arena at any token-grid width. Pooling\nalso snapshots both source tokens while their BF16 values are already in\nregisters, avoiding a separate full-hidden blit and redundant source read. The\nsame entry kernel keeps each pooled row in threadgroup memory and emits the\nfirst reduced block's attention AdaLN, eliminating another global residual read.\nAt the restore boundary, the first full-resolution attention AdaLN is fused\ninto expansion: a 10.5 KiB threadgroup row avoids a global residual reread while\nstill writing the exact bypass needed by the following residual branch.\nOn a thermal-balanced 512x512x22, 19-forward IT M5 Max A/B this reduced denoise\ntime from 39.13 to 28.06 seconds (28.3%). Final video/audio latent relative L2\nwas 5.56%/15.14%. First/middle/last fox frames retained one clean muzzle,\ncoherent legs, and sharp fur; an independent surfer remained consistent with\none rider and board through the wave spray. It changes composition and is\ntherefore opt-in rather than the close-reference default.\n`H3_TOKEN_REDUCTION_BLOCKS`\n\ncan override the later `4:30`\n\ninterval;\n`H3_TOKEN_REDUCTION_EARLY=STEPS:END`\n\noverrides the early schedule and `0`\n\ndisables it. `H3_DISABLE_TOKEN_REDUCTION=1`\n\nprovides an in-context exact oracle.\n`H3_DISABLE_FUSED_TOKEN_POOL_ADALN=1`\n\nand\n`H3_DISABLE_FUSED_TOKEN_ADALN=1`\n\nindependently restore the two-kernel entry and\nexit boundaries for diagnosis.\nToken reduction composes cleanly with the validated `--layers 45 --reuse 2`\n\nsettings: on the same 512 benchmark it reduced that profile from 16.69 to\n12.60 seconds (24.5% marginal), and independent fox and surfer renders stayed\ncoherent. Do not combine it with both `--layers 40`\n\nand `--reuse 3`\n\n; that\n6.47-second experiment produced chromatic ringing and ghosted limbs despite\nacceptable latent norms.\n\n`--render-width`\n\nand `--render-height`\n\nrun the model and VAE on a lower\nsame-aspect internal canvas, then high-quality vImage-scale RGB frames to the\nrequested output size before callbacks, terminal display, and encoding. This is an\nexplicit quality/speed tradeoff: a measured 384-to-512 prompt render reduced\nM5 DiT time by 33% and video-VAE time by 18% while retaining a clean,\nrecognizable photorealistic result. Both values must be multiples of 32; the\nexact output canvas remains the default.\nFor square 512 output, 384 is the fast-quality point and 320 is the validated\naggressive point. The latter produced a coherent walking fox and repeated at\n8.02 seconds of DiT versus about 15.82 seconds natively. Native 256 uses the\nsame-cost spatial-RoPE adaptation described above; it remains a fast composition\npreview rather than a substitute for a 512- or 768-class final render.\nThe video VAE automatically chooses a 256-320 pixel spatial tile from the\nrequested canvas geometry, minimizing repeated overlap work while keeping peak\nstorage bounded. `H3_VAE_TILE_PIXELS=256`\n\nrestores the original conservative\ntile plan for close-reference diagnosis.\n\nOn M5-class GPUs, persistent transformer weights are mapped directly from their\nsafetensor shards instead of copied into anonymous shared buffers. This keeps\nthe 37 GiB model file-backed/reclaimable and slightly improves total transformer\ntime; M3 uses the faster copied-buffer path. `H3_ZERO_COPY_WEIGHTS=0`\n\ndisables\nthe M5 selection for diagnostics.\nThe streamed Qwen text encoder preallocates a small ring of future layer\nbuffers and fills them on eight I/O workers while Metal executes the current\nlayer. The default ring depth is two layers on M3/older hardware and three on\nM5, where the target machine has 128 GiB. `H3_QWEN_PREFETCH=0`\n\nrestores the\nsingle-layer synchronous reference path; values 1-8 select the worker count,\nand `H3_QWEN_PREFETCH_DEPTH=1`\n\nthrough `6`\n\noverrides the ring depth.\n\nM5 GPUs automatically use native BF16 Metal 4/TensorOps for the DiT QKV and\nattention-output projections at sequence lengths up to 2,048. The compact\nMorton schedule routes Q/K/V directly into head-major attention inputs, avoids\nthree MPSGraph input transposes, and is byte-identical to the portable path. It\nimproves a complete 512x512 50-block forward by about 2% across repeated IT/US\nM5 Max runs. For 2,049-3,072 rows, including 864x480, two row-offset Morton\ndispatches preserve the efficient tile geometry and improve the complete\nforward by about 2% in balanced runs. Still larger sequences stay on MPSGraph.\n`H3_NAX=0`\n\ndisables TensorOps for exact A/B diagnosis. The selection is guarded\nat runtime and falls back to the unchanged portable library if compilation is\nunavailable.\n\n`H3_NAX=1`\n\nforces the broader native BF16 linear path. It passes the complete\n50-block MLX fixture, but remains opt-in: exact-shape microbenchmarks favor its\n128-row tile while full DiT runs currently favor MPSGraph scheduling. This\nkeeps a working NAX integration available for later quantized/fused kernels\nwithout making a benchmark regression the default.\n`H3_NAX=mlp`\n\nselects a more specialized Metal 4 path: paired FC1 gate/up\nTensorOps tiles apply SwiGLU in threadgroup memory and write only the\n14,336-wide activated intermediate, then FC2 also stays on TensorOps.\n`H3_DISABLE_NAX_MLP=1`\n\nkeeps the MPSGraph MLP in a context created this way for\nsame-process A/B testing. The path is deliberately opt-in because scheduling\ndepends on the OS GPU stack: the primary macOS 26.5.2 M5 Max gained 1.3-2.0%\nin isolated real-weight MLP runs but lost about 1-3% in a complete 50-block forward,\nwhile an otherwise identical macOS 26.5 M5 Max gained 1.4% in a same-context\nforward A/B. The resulting 50-block velocities were close (1.9% video and 2.4%\naudio relative L2), but not byte-identical.\n\nThe narrow DiT audio/video output heads convert their small released F32\nweights to BF16 once and use the Iris-derived 16x16 tiled linear directly on\nBF16 activations. At the production 320-render geometry, isolated paired-head\nmeasurements are 2.30x faster on M3 Max and 1.83x faster on M5 Max, with\nrelative L2 `8.64e-4`\n\n; the absolute M5 saving is about 0.6 ms per evaluated\nstep. Full fox and surfer sequences remained clean and measured 29.9/38.4 dB\nagainst the F32-head renders. `H3_DIT_F32_FINAL=1`\n\nrestores the close-reference\nhead and its extra activation buffers.\nThe F32 `96->5376`\n\nvideo and `32->5376`\n\naudio patch projections use a dedicated\n16x16 cooperative tile, retaining F32 weights, inputs and accumulation while\nrounding the tile result directly to BF16.\nPaired production-shape measurements are 1.77x faster on M3 and 1.62-1.78x\non M5; the complete generated RGB stream is byte-identical to the scalar path.\nFusing the final cast improves the 2835-row tile itself from 2.499 to 1.734 ms\non M3 and 1.555 to 1.186 ms on M5, and removes 38.27/59.66 MiB of F32 scratch\nat 512/864-class geometry. `H3_DISABLE_FUSED_PATCH_CAST=1`\n\nrestores the tiled\nF32 output plus standalone cast; `H3_SCALAR_PATCH=1`\n\nselects the scalar\ndiagnostic path.\nThe same tile binds its output directly into the packed hidden stream, removing\nthe BF16 media staging buffers and their blits. This saves another 19.13/29.83\nMiB and improves the 2835-row boundary from 1.847 to 1.730 ms on M3 and 1.282\nto 1.184 ms on M5. Contiguous T2VA uses byte offsets; FL2VA/Ref2VA use compact\ndestination-row maps so each modality remains one large dispatch. A complete\nsix-segment Ref2VA M5 ABBA remained byte-identical and improved 5.067 to 5.033\nseconds per measured forward pair. `H3_DISABLE_FUSED_PATCH_PACK=1`\n\nrestores the\nstaging buffers and packing blits.\n\nThe DiT core is split into two ordered Metal command buffers so GPU execution\nof the first part overlaps CPU encoding of the second. Thermal-balanced ABBA\nmeasurements select a 60%-depth split on M5 (30/50, 27/45, and 24/40), with\nroughly 0.5-1.8% wins; M3 automatically splits only the validated 30/50 case,\nwhich measured 1.2% faster, because 24/40 regressed there. The operation order\nand generated bytes are unchanged. `H3_DIT_COMMAND_BLOCKS=0`\n\nrestores one\ncommand buffer; values 1-50 override the split for further tuning.\nDiT activation buffers also follow their actual intra-block lifetimes: the QKV\nprojection arena is reused first for attention heads and then for the normalized\nMLP input, while the current attention-output arena becomes the MLP output after\nits branch has been consumed. This removes 61.25 MiB at 512-class geometry and\n99.63 MiB at 864-class geometry without changing dispatches or arithmetic.\n`H3_DISABLE_DIT_ACTIVATION_ALIAS=1`\n\nrestores separate diagnostic buffers.\nMPSGraph tensor-data wrappers for immutable DiT weights and biases are retained\nwith their resident buffers. This avoids rebuilding the same binding metadata\nfor every block and denoiser evaluation without copying tensor storage; measured\nABBA gains were 1.6% on M3 Max and 0.4-1.1% on M5 Max. Activation wrappers stay\ntransient because retaining them regressed the M5. The outputs remain\nbyte-identical, and `H3_DISABLE_GRAPH_DATA_CACHE=1`\n\nrestores transient wrappers\nfor all tensors.\nOn M3/older hardware, the four MPSGraph segments in each DiT block also reuse\none `MPSCommandBuffer`\n\nwrapper for their shared underlying Metal command buffer.\nRepeated thermal-balanced runs measured 1.0-1.6% faster on M3 Max; M5 measured\nneutral, so it retains fresh wrappers. `H3_REUSE_MPS_COMMAND=0`\n\nor `1`\n\noverrides\nthe automatic selection. Results are byte-identical.\nOn M5, the serving Euler sampler keeps its patch-packed F32 latents and cached\nBF16 velocities in Metal buffers. Each selected denoiser refresh is completed\nbefore the next is encoded, avoiding MPSGraph back-pressure while removing all\nintermediate latent/velocity readbacks and repacking. Two warm eight-run A/B\nsequences measured small 0.1% and 0.3% gains with byte-identical final latents;\nthe path also saves roughly 16 bytes of transient host state per video-latent\nelement (about 136 MB at the 768p shape). M3 and older GPUs retain the CPU\nsampler by default. `H3_CPU_SAMPLER=1`\n\nrestores it on M5;\n`H3_GPU_SAMPLER=1`\n\nselects the GPU-state path explicitly, and\n`H3_GPU_SAMPLER_WINDOW=0`\n\nenables the slower unbounded encode-ahead diagnostic.\n\nThe released checkpoint stores DiT QKV rows interleaved per attention head. Native Metal consumes that layout directly in the fused QK-normalization/RoPE kernel, avoiding a checkpoint transpose and extra RAM. The earlier identity interpretation was the cause of the noisy diagnostic outputs.\n\nThe public generation path decodes the joint audio latent with a streamed native\nBigVGAN/AudioVAE and writes synchronized H.264 plus 32 kHz stereo AAC. The native\nwaveform agrees with the corrected MLX oracle to relative L2 `6.94e-5`\n\n.\n`--first-frame`\n\n, `--last-frame`\n\n, and their combination use the released visual\nVAE encoder, Qwen3-VL vision tower and three-deepstack multimodal presentation,\n0.999 condition augmentation, and fixed condition rows in the native DiT. The\nfirst image is stretched to the target canvas; the last image is aspect-cover\nscaled and center cropped, matching the reference implementation. `--ref-image`\n\nselects the distinct Ref2VA transformer, preserves ordered ``\n\npresentation, and uses the released down-only aspect-preserving reference canvas.\n`--ref-silent-video`\n\nadditionally performs bounded 24 fps decoding, the visual\nVAE's causal `ceil(T/4)`\n\ncompression, two-frame Qwen sampling, and timestamped\n`