10Eros Max H3 Turbo-hybrid beta5 — learned proxy-activation quantization

[!WARNING] PlagueKind H3 MiniMax Cache users: A reported interaction with ComfyUI's AIMDO model compiler can cause aimdo memory compile error: could not start recording. Bypass that cache node if affected. See Format and compatibility below.

This repository contains quality-oriented, weight-only quantizations of TenStrip/10Eros-Max for ComfyUI. The current recommended deliverable is V8: the OUT- and edge-protected INT8/W6A8/W4A8 hybrid, available in final_int8w6w4_hybrid/.

10Eros_Max_h3_TURBO-hybrid_beta5_proxy_activation_int8w6w4_hybrid_v8_14_5gb_out_edge_protected_final.safetensors

Keep its accompanying .sha256 and .quantization.json files. The checksum identifies checkpoint bytes; the quantization report preserves assembly evidence. Neither is an inference dependency.

The previous V5 INT8/W4 release remains at repository root, with its original checksum, report and technical overview. The accepted experimental V7 W6/W4 hybrid remains in experimental_beta5_w6a8/. Both are preserved alternatives with their original evidence. V8's promotion reflects this campaign's mechanical validation and practical live A/B acceptance, not a guarantee of better results for every prompt, seed or workflow.

Earlier non-proxy learned releases remain under BasicLearned/, and the first accepted proxy-activation release remains under ProxyActivatedLearned_v1/. They are preserved for provenance and comparison and are not the current default.

A working learned W4A8 conversion of the non-Turbo beta5 transformer is also available as BasicLearned/10Eros_Max_h3_hybrid_beta5_w4a8_14gb_optimized.safetensors in this repository's BasicLearned/ folder. It is a completed and runtime-tested legacy alternative for users who specifically want the non-Turbo beta5 model, but it predates this proxy-activation campaign and is not the current recommended release in this repository.

This is a transformer-only derivative precision conversion. It does not contain the text encoder, video/audio VAE, vocoder, or workflow. Use the same compatible H3 ecosystem components as the source checkpoint. A sample two-pass workflow is included.

Content notice: the source repository is marked Not-For-All-Audiences. The same usage and content considerations apply to this derivative.

Current main artifact — V8 INT8/W6/W4 hybrid

Download the final V8 checkpoint.

All 50 attention-output matrices have a W6-or-INT8 precision floor. Blocks 0–1/47–49 require W6 g16 or INT8 for every main matrix; blocks 44–46 require W6 g32 or INT8. Other matrices are allocated between admitted W4, W6 and INT8 candidates. The final file has no W4 g8 even though that format was considered.

V8 reuses the source-matched V7 capture and learned W4/W6 populations and adds custom activation-informed INT8 learning across all 200 main matrices. Its allocation uses role-normalized reconstruction error, mean-one matrix-size weighting with exponent 0.75, and 4x protected-block weighting. It gives INT8 no global format bonus. This balances local error reduction against stored bytes while enforcing the OUT and edge precision floors.

The final checkpoint passed candidate/layout, preserved-byte, exact-size and integrity validation, then loaded and generated in live ComfyUI tests. The operator judged it the best checkpoint tested in this campaign. A matched 8x-protection control retained a few details better but introduced noticeable artifacts absent in the final 4x version; its small gains did not outweigh those losses. These are practical A/B observations, not a controlled perceptual benchmark or a promise of universal source fidelity.

V8 requires native H3, W4A8, W6A8 and INT8 ConvRot support in ComfyUI and compatible Comfy Kitchen dependencies. Its native descriptors are asym_w4a8_int8, w6a8_int8 and int8_tensorwise. Update ComfyUI and its dependencies before use; an installation supporting V5 alone may lack W6A8 support. No activation data or offline learning is required at inference.

For capture, error objectives, all three learners, allocation and acceptance, see the separate V8 proxy-activation technical overview. The folder also includes the quantization report and checksum.

Checkpoint size is a decimal file size, not total runtime VRAM. Runtime memory and speed depend on the workflow, other resident models and offloading. Controlled generation timing and VRAM traces were not recorded for V8.

Previous main artifact — V5

The final map covers the four large projection matrices in each of the 50 main transformer blocks:

Blocks 0--1 and 47--49 receive exact admitted INT8-p20 protection. Blocks 44--46 receive an additional W4-g8-or-better floor.

“Proxy-activation-calibrated” describes the offline W4 candidate learning, code refinement, scoring, and selection. The published W4 tensors remain the standard native weight-only asym_w4a8_int8 format and require no calibration activations at inference time.

V5: what “proxy activation” means

This release uses activations captured from a representative beta5 proxy checkpoint to estimate how quantization error propagates through real H3 input distributions. Calibration and prompt-disjoint development trajectories guide the offline W4 rounding refinements, selective learned INT8 candidates, safety guards, and final byte-budget allocation. The distributed checkpoint remains a weight-only W4A8/INT8 model; it does not require activation data at inference.

W4 construction has two stages. The activation-aware base learner first optimizes discrete codes and stored scales and can update the native-format 16-entry codebook through an exact-tested quadratic proposal followed by code reassignment. The final cumulative v5.7 refinement then freezes those already learned same-release scales and codebooks while combining exact bounded micro-greedy code changes with binary AdamW relaxation in both cumulative orders. A final population-wide, benefit-dependent microgreedy polish then evaluates up to eight adjacent-code passes for the W4-g16/g8 families that can participate in the protected allocation. This polish occurs before allocation; the assembler only materializes already validated payloads. A separate control panel found no better retained candidate from another joint-scale or codebook- polish pass. INT8 candidates use row-wise ConvRot with fixed scales and learned top_p=0.20 rounding. Candidate selection uses calibration evidence; development trajectories are audit evidence rather than a selection input.

For the capture design, mathematical objectives, W4/INT8 optimization, allocation, validation infrastructure, and numerical results, see V5 proxy-activation technical overview.

Format and compatibility

The W4 representation is ComfyUI-native asym_w4a8_int8: packed W4 codes, FP8 E4M3 relative group scales, FP32 channel scales, a per-matrix 16-entry FP32 codebook, and ConvRot group 256. INT8 matrices use standard row-wise INT8 ConvRot with FP32 row scales and ConvRot group 256.

Use a current ComfyUI build with MiniMax H3 and native W4A8 ConvRot support, together with compatible Comfy Kitchen dependencies. Install the checkpoint in the diffusion-model location used by your H3 workflow and select it as the transformer/diffusion model.

Third-party cache compatibility: If PlagueKind's experimental H3 MiniMax Cache node triggers aimdo memory compile error: could not start recording, bypass that node. Restarting ComfyUI with --disable-comfy-compiler can be used as a diagnostic workaround, although it may reduce performance. This interaction does not by itself indicate checkpoint corruption.

Runtime memory depends on resolution, frame count, attention backend, LoRAs, the W4/W6/INT8 mixture, and other resident models. The 14.5 GB figure is the checkpoint's decimal file cap, not total runtime VRAM. INT8 weights use more resident bytes than packed W4 weights; near a hardware limit, a higher INT8 share can reduce headroom and trigger additional offloading even when the checkpoint remains within its disk-size cap.

V5 validation and operator evaluation

• [x] The source-bound v5.5/v5.7 W4 populations, the final balanced microgreedy W4 population, and both INT8 populations passed their authoritative validators.
• [x] The frozen 200-matrix map passed exact size, header, descriptor, candidate-digest, preserved-byte, resume, and SHA-256 validation.
• [x] The final checkpoint loaded and generated successfully in the target ComfyUI workflow.
• [x] In the completed focused audio/video A/B set, every tested output was equivalent or superior overall to the preceding proxy-activation variants. The operator judged video and audio quality very good and promoted v5 as the definitive beta5 release.
• [x] Known proxy-family caveats remain: slight behavior drift from source beta5 despite prompt compliance, and occasional slight unwanted-audio or speech hallucinations.
• [x] V5 appeared slightly slower and more memory-hungry than some predecessors; controlled timing, transfer, and VRAM traces were not recorded.

The final observations are practical operator checks, not a controlled or repository-reproducible benchmark. Do not generalize the perceived improvement to every prompt, seed, LoRA, resolution, or workflow.

Experimental V7 W6A8/W4A8 hybrid

The experimental release is the operator-accepted protected 14.30 GB variant of the V7 campaign. Its exact size is 14.317977065 decimal GB. V5 remains an accepted historical alternative; V8 is now the main deliverable.

Download the experimental V7 checkpoint.

experimental_beta5_w6a8/10Eros_Max_h3_TURBO-hybrid_beta5_proxy_activation_w6w4_hybrid_v7_14_25gb_fc1_b44-46_w6.safetensors

The filename retains the 14.25 GB base allocation and identifies the three FC1 upgrades; it does not mean that the final protected file is only 14.25 GB.

Blocks 0–1 and 47–49 use W6 g16. All four main matrices in blocks 44–46 use W6 g32, including FC1. The remaining matrices are allocated between W4 g16 and W6 g32. This replaces V5's INT8 edge protection with W6 and extends W6 protection across the complete late-block 44–46 band.

Serialized tensor counts are 200 INT8 packed weight tensors, 200 FP8 E4M3 scale tensors, 285 FP32 scale/codebook tensors, 200 UINT8 descriptors and 334 BF16 tensors. INT8 storage dtype here is the byte container for packed four-bit or six-bit weights; it does not mean that V7 contains INT8 ConvRot matrices.

V7 method and compatibility

V7 uses a same-source 72-trajectory proxy-activation capture, with 48 calibration and 24 prompt-disjoint development trajectories. W4 uses short joint code/scale/codebook AdamW learning and guarded discrete polishing. The released W4 population combines independently measured positive-penalty and zero-penalty results, retaining the better qualifying endpoint for each matrix/group. W6 learns rounding on a uniform six-bit grid, keeps native scales fixed, confirms shortlisted penalties with full learning, and then applies guarded polishing. Development trajectories remain audit-only.

W4 uses native asym_w4a8_int8; W6 uses native w6a8_int8, with packed six-bit weights, FP8 relative group scales and FP32 channel scales, without a learned codebook. Both use ConvRot group 256. No activation capture or learning is required at inference time.

V7 requires ComfyUI and Comfy Kitchen with native W6A8 support in addition to the H3/W4 support needed by V5. W6A8 support was added in Comfy Kitchen PR #191. The campaign's quantization runtime used Comfy Kitchen 0.2.36. Update both ComfyUI and its compatible dependencies before trying the experimental file; an installation that supports V5 does not necessarily support V7.

For the capture, objectives, learner changes, protection policy, allocation, measured results and acceptance limits, see the separate V7 proxy-activation technical overview. The V5 overview continues to describe V5 alone.

The uploaded quantization report and assembly progress sidecar are retained as assembly evidence. Neither is an inference dependency.

V7 validation and operator evaluation

All four V7 comparison checkpoints passed mechanical validation and loaded and generated successfully in the operator's live ComfyUI tests. Quality was extremely close across the 14.00, 14.25 and 14.50 GB budgets, with small perceived changes in the expected direction. The protected 14.317977 GB variant appeared slightly above even the ordinary 14.50 GB checkpoint and was accepted as the final V7 artifact on 2026-10-04.

Comparison with V5 is mixed. Even with matching prompts and seeds, the changed quantization causes scenes to evolve differently. Some examples favored V5 INT8/W4 and others V7 W6/W4. These practical observations do not establish that V7 is generally better than V5. No controlled generation-speed or runtime-VRAM measurements were recorded, and a smaller checkpoint does not guarantee faster inference or proportionally lower peak VRAM.

The source-model license, content notice, transformer-only scope and workflow-dependent behavior described above also apply to V7.

Technical provenance and license

The quantization method and retained evidence are documented in the companion quantization repository. The W4A8, W6A8 and INT8 ConvRot layouts are provided by Comfy Kitchen and follow ComfyUI's quantization format.

This is an experimental derivative precision conversion, not an official upstream checkpoint. No ownership of the source model, training data, or associated workflows is claimed. This conversion grants no additional rights. Comply with the source checkpoint's license, terms, and content/use requirements, along with those of the surrounding inference components.