BF16 and INT8 ConvRot variants of the H3 text/vision conditioning encoder,
containing language layers 0–49; and
generation-only tails containing layers 50–63, the final norm, and LM head
in BF16, INT8 ConvRot, and NVFP4/AWQ formats.
H3 consumes the unnormalized hidden state after language layer 49.
This checkpoint therefore includes the Qwen3-VL embedding, language layers
0–49, and the complete vision tower. It intentionally omits language layers
50–63, the final language norm, and the LM head.
350 learned row-wise INT8 ConvRot language matrices
ConvRot group size 256 for every learned language matrix
One simple tensorwise INT8 token embedding
551 tensors retained in BF16, including the complete vision tower and all
norms
351 FP32 weight scales and 351 ComfyUI quantization descriptors
Use BF16 when memory permits. The INT8 ConvRot build is intended for systems
where the 47.97 GiB conditioning encoder is too large.
Generation tails
Every tail contains Qwen3-VL language layers 50–63, the final language norm,
and LM head. A tail is not a standalone CLIP: it reuses the tokenizer,
embedding, vision tower, and layers 0–49 from the connected H3 conditioning
encoder.
Language layers 50–63, final language norm, and LM head
98 learned row-wise INT8 ConvRot matrices
One simple row-wise INT8 ConvRot LM head, evaluated in chunks by the node
ConvRot group size 256
57 tensors retained exactly in BF16
The tail does not duplicate the token embedding or vision tower. It is loaded
temporarily, then unloaded after generation while the connected conditioning
CLIP remains unchanged.
ComfyUI installation
Place the selected conditioning encoder and optional tail under:
ComfyUI/models/text_encoders/H3/
Select it in CLIPLoader with the H3-compatible text-encoder type. Use a
current ComfyUI checkout with its pinned comfy-kitchen dependency.
For prompt enhancement:
Load the 0–49 conditioning checkpoint with ComfyUI's standard
CLIPLoader using the H3-compatible text-encoder type.
Connect that CLIP to H3 Prompt Enhancer (optional CLIP tail).
Select the 50–63 tail in the node's clip_tail dropdown.
Send enhanced_prompt and the returned, unchanged clip to the normal
H3 guide node.
If the connected CLIP is already a complete generative model, leave
clip_tail at [none — connected CLIP is already complete]. The enhancer
then calls the connected CLIP's ordinary generate() path, without loading
or requiring this tail.
These are ComfyUI checkpoints, not a complete Transformers generation
repository.
Standalone text and vision-language generation
Install
ethanfel/ComfyUI-H3-Qwen3VL-TextGen
to use the H3 0–49 conditioning encoder plus any compatible tail in this
repository as a standalone, general-purpose local Qwen3-VL text and
vision-language generator. It does not require the separate H3 prompt-guide
node pack.
Install or link the TextGen repository under ComfyUI/custom_nodes.
Put the H3 conditioning encoder and selected generation_tail_50_63 file
under ComfyUI/models/text_encoders.
Load the conditioning encoder with ComfyUI's standard Load CLIP node.
Select the tail with H3 Qwen VL Generation Tail Loader.
Connect both outputs to H3 Qwen VL Generate Text (Standalone).
The standalone node supports editable system/user prompts, optional image
batches, deterministic or sampled decoding, and Qwen thinking mode. The base
CLIP is preserved; only the temporary tail is explicitly unloaded after
generation.
Runtime verification
The conditioning checkpoint and enhancer passed actual runtime tests:
Detected the H3-compatible text-encoder model class
Finite conditioning output: (1, 12, 5120)
Correct modality-token tags: (12,)
VRAM after encode: about 24.7 GiB allocated / 26.1 GiB reserved
Standard CLIPLoader loaded the conditioning model with exactly 50
language layers and no final norm or LM head.
The optional tail path generated a token through all 64 layers, returned the
exact same CLIP object, then left it at exactly 50 layers with no norm/head.
The returned CLIP successfully encoded H3 conditioning after the tail
was unloaded: finite (1, 4, 5120) output with token tags.
The no-tail path was tested with a complete Qwen3-VL-4B ComfyUI CLIP and
generated successfully without loading the H3 tail.
The NVFP4/AWQ tail was tested with the matching 0–49 NVFP4/AWQ encoder on an
NVIDIA RTX PRO 6000 Blackwell. It generated through all 64 language layers,
unloaded cleanly, left the base at exactly 50 layers, and produced finite
(1, 7, 5120) H3 conditioning afterward.
The local CUDA 12.8 PyTorch build used fallback operations because this
comfy-kitchen release recommends CUDA 13.0+ for its optimized kernels. The
encode nevertheless completed successfully. A current ComfyUI environment
with its recommended PyTorch build is preferred.
Both upstream BF16 shards were downloaded at that revision and verified
against their Hugging Face LFS SHA-256 values before packaging.
The source model card reports Heretic v1.2.0 ARA edits targeting
attn.o_proj in language layers 31–40. All of those edited layers are inside
H3's retained 0–49 range, so the uncensoring edits are present in this
checkpoint. The source reports 4/100 refusals versus 99/100 for the original,
KL divergence 0.0421, PIQA 92.87%, and MMLU 79.87%.
Abliteration reduces refusal behavior but does not guarantee that every
refusal or safety behavior is removed, and it may affect model quality.
Conversion
The H3 BF16 package was converted with
silveroxides/convert_to_quant
1.3.1. The successful build used AdamW AdaRound optimization with plateau
early stopping, not simple rounding:
The language block matrices use learned ConvRot. Only the token embedding uses
simple tensorwise INT8 because ComfyUI embedding lookup requires that layout.
The vision tower is retained exactly in BF16.
The generation tail was packaged from the same pinned source and converted
separately:
The 98 transformer matrices use learned AdamW ConvRot. The LM head uses
simple row-wise ConvRot so the enhancer can compute its 151,936 output rows in
small chunks and avoid a multi-gigabyte temporary dequantization peak.
Validation
The completed file passed structural validation of every tensor, dtype, shape,
scale, per-layer descriptor, and global quantization metadata entry. All 551
protected BF16 tensors were compared byte-for-byte against the packaged BF16
source and were unchanged.
The tail also passed exact structural validation: the retained 57 BF16
tensors (304,128 bytes) are byte-identical to the source; its 99 INT8 weights,
scales, descriptors, and global quantization metadata all match the declared
layout. Combining the base source topology (902 tensors) with the tail source
topology (156 tensors) reconstructs all 1,058 tensors of the full model with
no key collision.
