license: other license_name: minimax-h3-community-license-agreement license_link: https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE pipeline_tag: image-text-to-video base_model_relation: quantized base_model:

TenStrip/10Eros-Max tags:
minimax-h3
comfyui
w4a8
w6a8
int8
convrot
quantization
activation-aware
text-to-video
image-text-to-video
image-to-video
not-for-all-audiences

10Eros Max H3 Beta7 fix1 eight-step Turbo — activation-guided learned quantization

Beta7 fix1 HQ (18.73 GB) is the default and recommended release. It uses learned W6 g16 and INT8 ConvRot weights from the fix1 transformer published by TenStrip/10Eros-Max, retaining its embedded eight-step Turbo. Mechanical validation and operator ComfyUI quality acceptance are complete, recorded on 2026-11-02.

There will be no W4/W6 compact companion to fix1. This release primarily addresses higher-resolution output, generally outside the intended use of memory-constrained compact setups. The accepted original Beta7 compact (14.65 GB) remains available and usable for those setups. It uses the original Beta7 source and does not contain the fix1 changes. Its allocation, measurements and acceptance are documented in the original Beta7 technical overview.

In several ComfyUI generations, the operator reported excellent fix1 HQ quality and often preferred its output to the matching native INT8 HQ proxy, despite its smaller checkpoint. These are subjective observations from the tested workflows, without a controlled generated-video quality or inference-speed benchmark. HQ identifies the high-quality precision map; it does not establish parity with the approximately 40 GB BF16 source.

This is a personal hobby project aimed at making useful H3 output accessible on consumer hardware with a smaller weight storage and memory footprint. Upstream source notes own the model recipe and generation recommendations.

Choose a checkpoint

Release	Exact size	Learned matrix formats	Intended choice
Beta7 fix1 HQ	18,734,317,942 bytes / 18.7343 GB	90 W6 g16, 110 INT8	Recommended; current fix1 source
Original Beta7 compact	14,647,885,469 bytes / 14.6479 GB	85 W4 g16, 103 W6 g16, 12 INT8	Accepted smaller-memory alternative; original source

Sizes are decimal disk GB, not required VRAM. Working memory, activations, other model components and offloading affect inference memory use. Compact was accepted with slightly lower observed quality than its original HQ and lower observed memory use; it has not been benchmarked against fix1 HQ.

Fix1 HQ uses INT8 for every attention OUT and FC2 matrix and all four main matrices at blocks 0/1/47/48/49. QKV and FC1 at blocks 2–46 use W6 g16. All 332 non-target tensors remain byte-preserved in BF16, including fix1's expanded video/audio output heads. Output metadata declares their 32 PDD heads for standard ComfyUI loading; no ComfyUI modification is required.

This is transformer-only: the text encoder, video/audio VAEs and workflow are separate. H3's BigVGAN waveform decoder is integrated into the audio VAE.

Optional text encoders

The official INT8 ConvRot checkpoint remains a quality-first option when resources comfortably allow it. Its 27.14 GB footprint makes smaller alternatives worth considering when memory is limited.

My Current Compact 15.35 GB W4/W6 encoder offers a practical balance of quality and resource use, with strong prompt adherence observed in testing and better measured weight reconstruction than both official 4-bit alternatives, INT4 and NVFP4, in the same size class. See the text-encoder README for the precision maps, reconstruction comparisons, runtime feedback and the 17.93 GB Current HQ alternative.

What activation-guided learning adds

Read the fix1 technical overview for the objective, capture recipe, precision map, family-level reconstruction measurements, safeguards and comparison limits.

Offline proxy activations guide reconstruction of the original BF16 weights. The completed fix1 HQ population gives these calibration results against native quantization at each matrix's selected precision. NRMSE is shown as a fraction; lower is better.

Raw activation NRMSE (sampled input rows):

Format	Matrices	Native mean NRMSE	Learned mean NRMSE	Mean paired reduction
W6 g16	90	0.015448	0.006302	59.41%
INT8	110	0.008193	0.003059	63.25%

Rank32-plus-residual activation NRMSE:

Format	Matrices	Native mean NRMSE	Learned mean NRMSE	Mean paired reduction
W6 g16	90	0.016236	0.010646	34.56%
INT8	110	0.008292	0.005703	32.26%

Each reduction is the equal-matrix mean of paired percentage reductions, not the ratio of the two mean-error columns. Different format populations are not a matched precision comparison, and these are not perceptual-quality gains. All 200/200 final tensors improve calibration raw and rank-plus-residual error against native; 199/200 improve diagonal-weighted activation error. Mean ordinary weight NRMSE increases 3.97% for W6 and 5.14% for INT8 relative to native. This is a measured tradeoff within the secondary guards. Matching native development baselines were not saved, so these percentages must not be described as held-out improvements.

Open the full-resolution reconstruction table. The figure includes native-error baselines and compares original Beta7 (recorded as “Test3”) with fix1. Sources, prompts, capture orientations, attention settings and learning objectives differ; this is a descriptive comparison, not an isolated learner A/B.

One unified learner supplies screening, main optimization and polish. All stages target a 50/50 native-relative mixture of rank-plus-residual and raw activation reconstruction, with a separate weight-preservation penalty. It uses fresh 96-step probes on approximately 1/16 of matrix chunks, adaptive same-family penalties and a second disjoint 1/16 refinement of the best three admitted candidates. Fresh full-matrix zero64 and screened-winner96 runs compete before guarded discrete polish. A positive probe favoring weight reconstruction supplies the polish penalty; native remains the fallback if no full run qualifies.

Fresh fix1 captures use 48 calibration trajectories from 12 prompts and 24 development trajectories from 8 disjoint prompts, balanced between euler/simple and er_sde/sgm_uniform. The revised prompts mix 640×480 and 480×640, with 226 frames at 24 fps, eight steps and video/audio shifts 12/3. Capture uses native VSA at 90% sparsity (10% kept) with blocks 0/1 dense, and comfy kitchen attention for dense routes. Development is audit-only and never guides learning or selection.

The proxy is a matching 22.94 GB native INT8 HQ companion, built from fix1: 180 interior matrices INT8, with edge blocks 0/1/47/48/49 and all non-target tensors in source BF16. It supplies activation inputs; the original BF16 weights remain the reconstruction targets. Proxy capture made this practical on an RTX 4080 with 16 GB VRAM and 32 GB system RAM. These inputs can differ from the BF16 model's; this is not exact BF16-teacher capture. No activation cache is needed at inference.

Quality observations and validation

The fix1 HQ checkpoint passed mechanical validation of selected candidates, source-preserved tensors, serialized layout, quantization metadata and output integrity. The operator then generated several videos in standard ComfyUI and accepted runtime behavior and perceived quality, often preferring it to the matching native INT8 HQ companion. These observations do not quantify VRAM, throughput or superiority over BF16, nor isolate the mixed objective's effect.

Fix1 HQ is 53.47% smaller on disk than its BF16 source, and 18.32% smaller than its native INT8 HQ proxy. Storage savings are not measured VRAM savings. The capture canvas describes the calibration envelope, not a generation-resolution limit. Follow the upstream source guidance and the supplied workflow notes for inference settings and limitations.

Use in ComfyUI

Place the selected .safetensors in the diffusion-model/transformer folder used by your H3 workflow and select it in the compatible loader. Use a ComfyUI installation with native W4A8, W6A8 and INT8 ConvRot support. W6 support is documented in Comfy Kitchen PR #191 and ComfyUI PR #16483. Use ComfyUI v0.38.0 or later for the provided checkpoints. Use the source's compatible H3 text encoder, video/audio VAEs and eight-step Turbo workflow. Capture settings describe the calibration envelope; 640×480 is not a restriction on generation resolution.

Keep the .sha256 and .quantization.json companions for integrity and provenance. They are not inference dependencies. These models are already quantized; use their native formats rather than applying another quantization.

The workflows folder provides a customizable ComfyUI workflow for two-pass H3 video/audio generation with learned latent upscaling and refinement. Additional features include first-pass Preview with cache reuse, optional sparse attention, continuation mode with video/audio guidance and stitched export, RTX super-resolution, and Normal-mode RTX frame generation for high frame rate videos. H3 generation and continuation remain at the native 24 FPS in this workflow. Optional output grain and selectable H.264 encoder presets use Lokken VideoTools, available through ComfyUI-Manager.

The latest workflow is at the folder's top level; previous versions are archived under old/. Check model selections and local paths against your installation, and follow the usage notes included in the workflow. The configured components are recommended.

Exact identities

Fix1 source: 10Eros_Max_h3_8stepTURBO_beta7_fix1.safetensors
40,265,608,072 bytes
SHA-256: 8a670bf9962e7379946f10bede480d3e603d0d1ed2ab02c22d36398f4c796a29

10Eros_Max_h3_8stepTURBO_beta7_fix1_activation_learned_HQ.safetensors
18,734,317,942 bytes
SHA-256: 958717347e004fd2b9dc08966d8b7464e1b6004ceae263b5de730a4a1bcb999b

Original Beta7 compact (different source; no fix1 changes)
10Eros_Max_h3_8stepTURBO_beta7_activation_learned_compact_w4w6int8_14.65gb.safetensors
14,647,885,469 bytes
SHA-256: 64bf10e2ef650c2484c60a0eb3bc495bba707aadc50281361ad0dd8bba7c7b1e

The original compact's source identity and allocation data remain in the original Beta7 overview.

Legacy Beta5 Turbo HQ references

Two native, unlearned Beta5 Turbo HQ references are also available in beta5/Beta5_HQRef:

Checkpoint	Size	Main-matrix topology
Native INT8 / W6 g16 HQ	18.69 GB	Blocks 0-1 and 47-49: all four families INT8. Blocks 2-46: QKV/FC1 W6 g16, OUT/FC2 INT8.
Native INT8 with BF16 edges HQ	22.89 GB	Blocks 0-1 and 47-49: all four families preserved BF16. Blocks 2-46: all four families INT8.

Both retain the Beta5 source's embedded Turbo and copy all 334 non-target tensors exactly in BF16. INT8 uses native ConvRot quantization; W6 g16 uses native quantization with scale search. Neither uses activation captures or learned weights, so the Beta7 learning gains above do not apply to them. They are transformer-only references for the legacy Beta5 source.

Previous releases

Original Beta7 HQ and its sidecars are retained in the Beta7 archive, alongside the original technical overview. The original compact remains at the repository root.

The accepted Beta6 releases remain historical references:

Older Beta5 folders retain previously released checkpoints and evidence:

They are different source campaigns. Beta7 gains, capture envelope and maps must not be attributed to those artifacts.

Attribution and evidence

Source model by TenStrip, based on MiniMax H3. Licensing follows the MiniMax H3 Community License Agreement and the upstream source's terms. The source is designated Not-For-All-Audiences; that designation also applies to this derivative.

The quantization project contains source-bound learner code, frozen policies, allocation maps and acceptance evidence. See the fix1 technical overview for the complete explanation of what was learned, measured and accepted.

TenEros_Max_optimized_learned