MOSS-Image Text Encoder — Heretic — DHQ GGUF

GGUF builds of pottokao/MOSS-Image-Text-Encoder-Heretic (decensored Qwen3.5-9B) for MOSS-Image.

One file does two jobs: it rewrites a short idea into a full MOSS-Image caption, then encodes that caption for the DiT. Weights are loaded once. A Q4-class file fits an 8 GB card.

The quantization is DHQ (DiT-aware quantization): every tensor's format is chosen from measured damage on two rulers at once — the error after MOSS-Image's own text stage (what the DiT sees) and the KL of the rewrite — instead of llama.cpp's fixed per-type rules. At the same file size, every DHQ build beats the standard llama.cpp quant on both rulers.

How to use

Use the Dual-Head Dragon ComfyUI node: https://github.com/pottokao-dotcom/ComfyUI-DualHead-Dragon (MOSS-Image support: products/moss.py).

1. Put a .gguf from this repo (and mmproj-moss-te-f16.gguf if you want image input) in ComfyUI/models/text_encoders/.
2. Dual-Head Dragon Loader (GGUF) → gguf = this file, type = moss_image (auto also detects it).
3. Connect its CLIP output to a normal CLIP Text Encode, or to Dual-Head Rewrite first to expand a short idea (the rewrite runs with thinking off, as a prompt enhancer does).
4. DiT / VAE as usual: moss_image_sft_* (or a Turbo / DMD build) and flux2-vae_small from Comfy-Org/MOSS-Image.

The loader reads the last layer after the final RMSNorm with MOSS-Image's own chat template (open think block), exactly what the official pipeline encodes.

Loaders other than the Dual-Head Dragon node: as of this release, ComfyUI-GGUF does not map Qwen3.5 GGUF tensor names, so it cannot load these files yet. The K builds contain K-quants only (no IQ tensors, checked file by file), so they are ready for loaders that do. The IQ builds are for llama.cpp-based engines only (ComfyUI-GGUF dequantizes IQ types on the CPU).

Files

Error = relative L2 error after MOSS-Image's text stage, on 36 held-out prompts, run through the real llama.cpp engine (lower is better). "vs heretic" isolates quantization; "vs official" is what the DiT trained on (it includes the heretic shift itself, 13.5% for the unquantized model). Rewrite KL = llama-perplexity KL vs the BF16 model on 80 of its own rewrites.

Dual-head (rewrite + encode), K-quants only

Dual-head, IQ (llama.cpp engines only)

On this model the IQ builds gain little: the K 4.7 GB file beats IQ 4.5 GB on both rulers. Part of the reason is known: the simulated IQ errors used for allocation do not always match what llama.cpp writes (e.g. forcing layer 0 linear_attn.out_proj from IQ3_S to Q8_0 in a stock IQ3_M file changes nothing, though the simulation flagged it), so the IQ allocation is less well informed than the K one. The files are still measured on real GGUFs, as for all rows.

Encoder only (no lm_head, no MTP head — cannot rewrite)

Removing lm_head and the MTP head changes nothing for encoding (identical error to the full BF16 model).

Vision tower

mmproj-moss-te-f16.gguf (0.92 GB, F16) — the Qwen3.5 vision tower, identical to the official one (Heretic did not touch it). Needed only for image input (looking at a reference image while rewriting, community edit LoRAs).

What the DiT sees

Same prompt and seed, MOSS-Image SFT Turbo int8 (8 steps, cfg 1), every encoder run through the Dual-Head Dragon node. The text-stage errors above barely show; the heretic encoder renders like the official one, and the signage is spelled right down to 3.9 GB.

Low-bit files (4.2 GB and below) keep images intact, but their rewrite KL is high: expect a weaker prompt enhancer.

How it was made

• Base: pottokao/MOSS-Image-Text-Encoder-Heretic → BF16 GGUF (llama.cpp, MTP head kept), imatrix from calibration_datav3 + 78 MOSS-style captions (PPL 7.29).
• Sensitivity: per-tensor, per-format Fisher probes on both rulers (24 prompts × 4 probes; 9 formats from Q2_K to Q8_0, including IQ3_XXS / IQ3_S / IQ4_XS). The DiT ruler is MOSS-Image's text_embedder (proj + 2 blocks), the same weights in SFT and Turbo.
• Allocation: start at the floor, buy upgrades by damage removed per byte until the budget (the same-size llama.cpp file) is spent. Rewrite and encoder are weighted equally in the dual-head builds; encoder only for the encoder-only builds.
• The single most sensitive tensor is layer 0 linear_attn.out_proj. llama.cpp's rules leave it at Q4; moving only that tensor to Q8_0 (8 MB more) lowers the stock Q4_K_M error from 9.1% to 8.0% (measured). DHQ does this and re-allocates the other 247 tensors as well, which is where most of its gain comes from. (The linearized probes overstate this tensor's share — they put it at 65% of the Q4 damage — so treat per-tensor shares as a ranking, not as exact sizes.)
• recipes/ holds every file's per-tensor type map and the imatrix, so each file can be rebuilt bit for bit with llama-quantize --imatrix recipes/imatrix.gguf --tensor-type ... <bf16.gguf> <out.gguf> <base type>. SHA256SUMS lists the file hashes.

Use responsibly

Adults only. Do not depict minors or age-ambiguous people, non-consensual acts, or real people without consent.

License

Apache-2.0, same as Qwen3.5-9B.