MiniMax-H3 · fl2va × ref2va Hybrid b25-49 (W6A8, ~14.88 GiB)

A 6-bit weight / 8-bit activation build of the community-standard b25-49 hybrid. Single file, drop-in for ComfyUI's stock Load Diffusion Model.

minimax_h3_hybrid_fl2va_ref2va_b25-49_w6a8.safetensors   15,983,746,636 bytes (14.88 GiB)   1151 tensors

Between the two extreme choices: W4A8 (11.68 GiB, coarser 4-bit) and INT8 (20.97 GB, biggest). Per Comfy-Org's own comfy-kitchen notes, the 6-bit path trades 1.5× the bytes of 4-bit for ~3× lower weight error — it is the accuracy-leaning middle of the H3 quantization family.

What b25-49 means

MiniMax-H3 ships two checkpoints with identical architecture and tensor layout:

• fl2va — first/last-keyframe conditioning, noticeably better visual/audio quality, no reference slots.
• ref2va — adds image/video/audio reference conditioning, slightly lower base quality.

A tensor-by-tensor comparison shows the two are effectively identical everywhere except the per-block adaln_proj weights (the AdaLN modulation projections that route conditioning signal into each block). The bX-49 hybrids keep fl2va as the base and take adaln_proj from ref2va for blocks X–49 only.

b25-49 = blocks 25–49 from ref2va (the community default balance point). b30-49 leans quality, b20-49/b15-49 lean reference adherence.

How this file was built — no retraining, no re-quantization

A tensor transplant between already-quantized files, so no additional numerical error is introduced:

The file header (120,112 B) is byte-identical to the skeleton, so every per-layer comfy_quant descriptor is untouched and the data region keeps a 1:1 offset mapping. Same recipe as the W4A8 sibling below — only the quantization level of the two sources differs.

Verification

• Header byte-identical to the skeleton; file size unchanged (15,983,746,636 B); data region tail-aligned.
• All 100 per-block adaln_proj tensors checked one by one: blocks 0–24 identical to the skeleton, blocks 25–49 identical to the donor, 0 anomalies.
• Non-quantized tensors sampled across the file: identical to the skeleton (except the 50 intentionally replaced adaln_proj tensors).
• Cross-quantization check (the strongest evidence that the hybrid semantics are right): adaln_proj values are quantization-independent, so this file's blocks 0–24 must equal the INT8 b25-49 hybrid's blocks 0–24, and its blocks 25–49 must equal that file's blocks 25–49 — both hold, byte for byte, even though the two files carry different quantization. The same identity was previously verified against Kijai's W4A8 conversions.

Quantization profile

• 201 layers {"format": "w6a8_int8", "group_size": 32, "convrot": true, "convrot_groupsize": 256} = 200 block linears (attn.qkv_proj, attn.out_proj, mlp.fc1, mlp.fc2 × 50 blocks) + condition_proj. Weight is int8 [N, 3K/4] holding uniform 6-bit codes (code − 32, no codebook), plus an fp8 per-group weight_s_rel and a fp32 per-channel weight_s_channel.
• 8 token_refiner linears: int8_tensorwise.
• Norms, adaln_proj, patch projections and output layers stay at BF16/F16/F32.
• Configuration lives in per-tensor .comfy_quant blobs (this file has no __metadata__ block).

Hardware note

w6a8_int8 shares the layout class and kernels with asym_w4a8_int8 (AsymW4A8Int8Layout, MIN_SM_VERSION = (8, 0)), so the native fast path requires compute capability ≥ 8.0 (Ampere or newer). On older cards it still loads and computes correctly, but runs through the portable fallback — measured on a GTX 1660 SUPER (sm_75), the W4A8 fallback was ~2.6× slower per layer than INT8 convrot. If you are on Turing / GTX 16-series, the INT8 build is the faster choice there.

Usage (ComfyUI)

1. Put the file in ComfyUI/models/diffusion_models/.
2. Load with the stock Load Diffusion Model node (requires a ComfyUI version with ConvRot INT8 + W6A8 support; the log should show Detected mixed precision quantization).
3. The model routes as ref2va: keyframes via MiniMaxH3AddKeyframes, subject/scene/style references via MiniMaxH3ReferenceToVideo — both usable in the same generation. Plain text-to-video works as usual.
4. Use the usual H3 text encoder and video/audio VAEs.

Related files

• W4A8 build of the same hybrid: https://huggingface.co/binglingzhimeng/minimax_h3_hybrid_fl2va_ref2va_b25-49_w4a8_mixed
• INT8 build of the same hybrid: minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors (community, e.g. smhfacct / RunningHubAI packaging)

Credits & license

• Original weights: MiniMaxAI/MiniMax-H3 (FL2VA + REF2VA).
• ComfyUI packaging / quantized conversions: Comfy-Org/MiniMax-H3 and Kijai/MiniMax-H3-experimental.
• Hybrid idea (adaln-only block splice): smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models.

No training, fine-tuning or gradient optimization was involved — tensor-level weight selection on top of existing quantized conversions. The merge/conversion is released under Apache-2.0; the underlying MiniMax H3 weights remain subject to the MiniMax H3 Community License.

---

中文摘要

这是社区默认档 b25-49 混合模型的 W6A8（6bit 权重 / 8bit 激活）版本（14.88 GiB，1151 张量）， 位于 W4A8（11.68 GiB，4bit）与 INT8（20.97 GB）之间：按 Comfy-Org 自家 comfy-kitchen 的说明， 6bit 用 1.5 倍的体积换来约 3 倍更低的权重误差，是精度取向的中间档。

做法与已发布的 w4a8 版完全同构：以 fl2va_pruned_w6a8 为骨架逐字节复制，只把 blocks.25–49 的 adaln_proj（50 个张量、43,545,600 字节）覆盖为 ref2va_pruned_w6a8 的值，不重新量化， header 逐字节不变（量化配置未改动）。

验证：100 个 adaln 张量逐个核对（0–24 来自骨架、25–49 来自供体，零异常）；抽样未量化张量与骨架一致； 并做了跨量化档交叉验证——adaln 取值与量化档位无关，本文件 0–24 与该 INT8 版 0–24、25–49 与该 INT8 版 25–49 均逐字节相同，证明 b25-49 的语义确实成立。

硬件提示：w6a8_int8 与 asym_w4a8_int8 共用同一 layout 与内核，原生加速需要 sm_80+（Ampere 及以上）； 老卡（如 GTX 16 系 / sm_75）能加载也能算对，但走软件回退，那里 INT8 版更快。