Add non-Turbo beta5 policy-close W4A8 quant

10Eros_Max_h3_hybrid_beta5_policy_close_w4a8.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:81684c58c42455d1790071d96273f1164d7b7c35cb1f50bbebe77ef1ba2a6d75
+size 13997644777

10Eros_Max_h3_hybrid_beta5_policy_close_w4a8.sha256

@@ -0,0 +1 @@
+81684c58c42455d1790071d96273f1164d7b7c35cb1f50bbebe77ef1ba2a6d75  10Eros_Max_h3_hybrid_beta5_policy_close_w4a8.safetensors

README.md

@@ -56,4 +56,47 @@ Due to a ton of confusion I've let Claude compile a full MD on the grafting, inc
 
 This is an evolving project subject to future fixes in H3 training. Training is clearly problematic, and my branch will rely on Sulphur H3 tunes. But I didn't want to wait for all that so I took the data from older models, grafting it in where the model needs it, grafting it to attn layers at a low level that doesn't disturb H3's visual or audio output quality. That's essentialy it.
 
-The standard H3 community license applies. Because this release now carries transferred character from LTX 2.3, Wan 2.2, and Krea 2, the community licenses for those source models apply as well to the portions of character that came from each.
+The standard H3 community license applies. Because this release now carries transferred character from LTX 2.3, Wan 2.2, and Krea 2, the community licenses for those source models apply as well to the portions of character that came from each.
+
+## Non-Turbo policy-close W4A8 quantization
+
+This PR adds a size-focused ComfyUI-native quantization of the full non-Turbo
+beta5 checkpoint:
+
+`10Eros_Max_h3_hybrid_beta5_policy_close_w4a8.safetensors`
+
+It is intended for the non-Turbo workflow used with the Turbo LoRA, including
+the H3 Power LoRA setup. It follows the observable compression layout of
+LokkenJP's optimized 14 GB Turbo quant, while recalculating all quantized
+payloads from the non-Turbo weights.
+
+### Quantization layout
+
+- 180 interior projection matrices: native `asym_w4a8_int8` ConvRot.
+- 20 edge projection matrices: row-wise ConvRot INT8.
+- ConvRot group size: 256.
+- W4 group allocation: 82×g16, 53×g8, 45×g4.
+- 334 remaining tensors retained in BF16.
+- Transformer-only checkpoint; text encoder, VAEs, vocoder, and workflow are
+  supplied separately.
+
+### Relationship to earlier quantizations
+
+| Variant | Source | W4A8 | INT8 | BF16 | Approx. size |
+|---|---|---:|---:|---:|---:|
+| Previous graft-preserving | Non-Turbo beta5 | 64 | 120 | 350 | ~19.2 GB |
+| Lokken optimized | Turbo beta5 | 180 | 20 learned | 334 | ~14 GB |
+| This PR | Non-Turbo beta5 | 180 | 20 standard ConvRot | 334 | ~14 GB |
+
+### Limitations
+
+This is recipe-close and size-matched. The 20 INT8 edge tensors use standard
+reproducible row-wise ConvRot rather than Lokken's private learned-rounding
+optimizer. Exact quality or prompt-adherence parity is not claimed. The
+checkpoint passed safetensors structure, native metadata, companion-tensor,
+and CUDA dequantization checks; no controlled perceptual A/B benchmark is
+claimed for this conversion.
+
+The source repository is marked Not-For-All-Audiences and remains subject to
+the applicable MiniMax-H3, LTX 2.3, Wan 2.2, and Krea 2 community-license
+conditions.