Ideogram-4-Text-Encoder-Heretic-DHQ-GGUF

Ideogram 4 Text Encoder (Heretic Qwen3-VL-8B) — DHQ GGUF

DHQ = Dual-Head Quant. The heretic (abliterated) Qwen3-VL-8B from Qwen-Image-2.1-Text-Encoder-Heretic, quantized so that every tensor gets the precision that matters to what the Ideogram 4 DiT actually reads (13 interleaved hidden layers + the final-norm output), measured per tensor with gradients through the DiT's text input stage. Each file is exactly the size of the standard llama.cpp file of the same name.

Tuned for Ideogram 4 only. For Qwen-Image-2.1 use Qwen-Image-2.1-Text-Encoder-Heretic-DHQ-GGUF.

About the DHQ series

DHQ files are quantized for use as a diffusion model's text encoder: each tensor's precision is chosen by how much its quantization error changes what the target DiT reads (measured through the DiT's own text-input stage), not by next-token loss. They stay normal LLM GGUFs, so the same file can still write or rewrite prompts: where a model is also used as a prompt enhancer (Z-Image-Engineer-V6, the heretic Qwen3-VL-8B), every file is also checked for that (the PE column) and the tables show where it gets worse.

DHQ repos: Qwen-Image-2.1 · Ideogram 4 · Z-Image-Engineer-V6 · Qwen3-4B (Z-Image / FLUX.2 klein 4B)

We also built Dual-Head Dragon, a ComfyUI node that runs one GGUF as both the text encoder and the prompt enhancer (llama.cpp inside). It is still in testing and not released yet, including its LoRA support (separate LoRAs for the encoding and the rewriting head) and DFlash speculative decoding for faster rewriting. The files here work without it (see How to load).

Results

Same heretic bf16 source, same size. Standard = llama.cpp default rules + the same imatrix. TE = relative error of the text conditioning after Ideogram 4's own text input stage, on 36 held-out prompts never used to build or pick anything (lower is better). PE = mean KL divergence of next-token predictions vs bf16 on 80 prompt-rewrite samples, i.e. how well the same file still works as a prompt enhancer (lower is better).

IQ4_XS, Q3_K_M, IQ3_M and IQ2_M are being re-tuned so prompt writing is at least as good as the standard file; they are not uploaded yet (the rows above are the first version).

Where PE is not bold, the DHQ file writes prompts worse than the standard file — clearly so for IQ4_XS, Q3_K_M, IQ3_M and IQ2_M. Ideogram 4 reads 13 interleaved middle layers, and the bits it needs are not the ones prompt writing needs. If you also use the text encoder to rewrite prompts at those sizes, the standard file is the safer pick.

K or IQ?

Each size step comes in two series built the same way, with different llama.cpp formats:

• K files (Q*_K*): simple block formats. Fast to decode everywhere — GPU, CPU, Metal, Vulkan — and on the GPU in ComfyUI-GGUF.
• IQ files (IQ*): codebook formats. For a similar error the file is smaller (compare the rows above), but decoding is heavier: fine on llama.cpp GPU backends, slower on CPU, and ComfyUI-GGUF decodes IQ3/IQ2 tensors on the CPU (tens of seconds per prompt). Our IQ4_XS files also contain some IQ3 (and in some repos IQ2) tensors, so they are slow in ComfyUI-GGUF too.

ComfyUI-GGUF → use a K file. llama.cpp (or Dual-Head Dragon) → either; IQ if you want the smallest file for a given error.

How to load

Dual-Head Dragon / llama.cpp: these are standard qwen3vl GGUFs; the scores above were measured this way.

ComfyUI: needs ComfyUI-GGUF and the add-on ComfyUI-GGUF-Qwen3VL-TE (commit 1f52b00 or later). Put the GGUF and mmproj-ideogram4-te-heretic-dhq-f16.gguf (this repo) in models/text_encoders/ without renaming them (the mmproj is found by file-name prefix), then CLIPLoaderGGUF with type ideogram4.

[!WARNING] Without the add-on (or with an older version) ComfyUI does not see the vision tower, treats the file as plain Qwen3-8B and builds the wrong text encoder without an error: ideogram4 falls back to its text-only encoder (per-token cos 0.84–0.87 vs bf16). With the add-on the conditioning matches the bf16 safetensors encoder (cos 0.9999, checked with a Q8_0 GGUF).

Reproducibility

recipes/ has one recipe per file: source repo and revision, llama.cpp commit, importance matrix (sha256), and the type of every tensor as built. These files were built on aarch64 (GB10); rebuilding a recipe on aarch64 gives a bit-identical file (checked for every uploaded file). On x86_64 llama.cpp rounds a few blocks differently.

License

Apache-2.0, same as the base weights (Qwen3-VL-8B-Instruct, Alibaba Cloud / Qwen). See LICENSE and NOTICE.