Qwen3-VL-32B H3 ComfyUI Generation Tails and Overlay

This repository contains generation-only tails for Qwen3-VL-32B H3 conditioning encoders. Each tail supplies language layers 50–63, the final language normalization layer, and the LM head.

It also contains an experimental generation-only LoRA overlay that can modify all 64 language layers temporarily during prompt generation. The overlay does not modify the H3 encoder used afterward for video conditioning.

These files are not standalone CLIP checkpoints. They reuse the tokenizer, token embedding, vision tower, and language layers 0–49 from a connected H3 conditioning encoder.

Files

File	Purpose	Format	Size
`qwen3vl_32b_h3_instruct_generation_tail_50_63_int8_convrot.safetensors`	Official Qwen layers 50–63, norm, and LM head	INT8 ConvRot	7,609,128,659 bytes
`qwen3vl_32b_h3_instruct_generation_tail_50_63_bf16.safetensors`	Official Qwen layers 50–63, norm, and LM head	BF16	15,208,606,744 bytes
`qwen3vl_32b_h3_instruct_generation_tail_50_63_nvfp4_awq.safetensors`	Official Qwen layers 50–63, norm, and LM head	NVFP4/AWQ with BF16 norm and LM head	5,396,902,102 bytes
`minimax_h3_qwen3vl32b__prompt_generation_overlay__polaris_r16_plus_heretic_v2.safetensors`	Experimental Polaris + Heretic-v2 prompt-generation overlay, layers 0–63	BF16 LoRA factors	273,288,216 bytes

The Heretic method used here changes projections in language layers 31–40. Those layers are part of H3's connected 0–49 encoder, not the 50–63 tail. Consequently, separate Heretic-labelled tails would contain no abliteration or uncensoring changes and are not needed.

Standalone text and vision-language generation

Install ethanfel/ComfyUI-H3-Qwen3VL-TextGen to reconnect an H3 conditioning encoder and one of these tails as a general-purpose local Qwen3-VL text and vision-language generator.

Load CLIP (H3 0–49 encoder) ── clip ──────────┐
                                              ├─ H3 Qwen VL Generate Text (Standalone)
H3 Qwen VL Generation Tail Loader ─ tail_clip ┘
Optional IMAGE batch ───────────── image ─────┘

The standalone node supports editable system and user prompts, optional image batches, deterministic or sampled decoding, Qwen thinking mode, and an optional generation overlay. The connected base CLIP is preserved; overlay weights are applied only to temporary generation modules and the temporary tail is unloaded after generation.

Installation

Place the selected tail under ComfyUI's text-encoder model directory, for example:

ComfyUI/models/text_encoders/H3/generation_tails/

The filename must contain generation_tail_50_63 so the dedicated loader can discover it. Restart or refresh ComfyUI after installing a new file.

Place the optional overlay anywhere under:

ComfyUI/models/loras/

Its filename contains h3 and generation_overlay, allowing version 0.3.0 or newer of ComfyUI-H3-Qwen3VL-TextGen to list it in the standalone generation node. Start with overlay strength 1.0. It is a text-generation LoRA, not an H3 diffusion-model LoRA.

Then:

Load the H3 0–49 conditioning encoder with ComfyUI's standard Load CLIP node and its H3-compatible encoder type.
Select the tail with H3 Qwen VL Generation Tail Loader.
Connect clip and tail_clip to H3 Qwen VL Generate Text (Standalone).
Optionally select the experimental overlay in generation_overlay.

The same tail artifacts can also be used by a compatible H3 prompt-enhancer workflow.

Formats

BF16

The BF16 tails retain all 156 source tensors at full precision.

INT8 ConvRot

The INT8 tails contain 98 learned row-wise ConvRot transformer matrices and a chunk-compatible INT8 LM head. Norms and other protected tensors remain BF16. The ConvRot group size is 256.

NVFP4/AWQ

The NVFP4/AWQ tail contains 98 native ComfyUI NVFP4/AWQ transformer matrices. Its final norm and 151,936 × 5,120 LM head remain BF16 so the tail loader can evaluate the vocabulary head through its ordinary full-precision path. The file contains 576 tensors and is 5.03 GiB.

Runtime verification

The NVFP4/AWQ tail was tested with its matching 0–49 NVFP4/AWQ encoder in a current ComfyUI container on an NVIDIA RTX PRO 6000 Blackwell. A generation step traversed all 64 language layers, the temporary tail unloaded cleanly, the connected base remained at exactly 50 layers, and the returned base CLIP then produced finite (1, 7, 5120) H3 conditioning.

The BF16 and INT8 tail paths were also exercised through all 64 language layers and verified to restore the connected 50-layer conditioning encoder after generation.

The experimental overlay was structurally validated as 448 adapters covering seven projections in every language layer 0–63: 340 rank-16 adapters and 108 rank-17 concatenated Polaris-plus-Heretic adapters. All factors are BF16 and finite. It has not received the same end-to-end runtime qualification as the three tails above.

Provenance

Official BF16 and INT8 tails:

repository: MiniMaxAI/MiniMax-H3
revision:   83db0c0efe6ef9824e0e194be110346c0a9542ed
architecture: Qwen3-VL-32B-Instruct

Official NVFP4/AWQ tail donor:

repository: cybermotaz/Qwen3-VL-32B-Instruct-NVFP4
revision:   dfcab1395094f56d3c78ba472158284def0fc278

Experimental generation overlay:

official base:       Qwen/Qwen3-VL-32B-Instruct
official revision:   0cfaf48183f594c314753d30a4c4974bc75f3ccb
Heretic-v2 source:   coder3101/Qwen3-VL-32B-Instruct-heretic-v2
Heretic-v2 revision: 6f13e040e1b95f36c439e8c9e3f7e143e0bf6d85
Polaris dataset:     TeichAI/polaris-alpha-1000x
dataset revision:    45053b0656d74d88cefeec606af2a5e733b2f3d7

The overlay is an experimental public-lineage reproduction. It is not the deleted Polaris-HI16 checkpoint and does not claim to reproduce its unpublished HI16 recipe.

See SHA256SUMS for every file's Hugging Face LFS checksum.

ethanfel/Qwen3-VL-32B-H3-ComfyUI-Generation-Tails