Gemma 3 12B IT QAT — Heretic-MOE v1

Work-in-progress release of Gemma 3 12B IT QAT variants produced with Heretic-MOE. The release is intended for two related uses:

1. a full language model for prompt rewriting and directing;
2. a native ComfyUI text encoder for LTX workflows.

The adaptive search, selected-model export, GGUF conversion, structural text-encoder validation, and release upload are complete. The final independent HF audit verified all 47 published artifacts by byte size and SHA-256.

Selected variants

The five finalists were re-evaluated together with fresh generations and a deeper perplexity pass. The requested 64 x 1024-token pass yielded 53 complete windows from the frozen built-in corpus, so all values below use the same actual 53 x 1024-token measurement. They form a separate comparison scale from the 24 x 512-token search metric.

For reference, the unmodified model measured SRG +0.00834, R-side 54.4%, and 12/136 keyword flags in the same recheck. SRG, R-side, keyword counts, and PPL are search diagnostics rather than semantic ground truth.

The search metrics are selection signals, not final behavioral claims. Final variants will also be evaluated with longer generations and blind semantic judging.

Full-model GGUF files

The Director GGUF files contain only the language model: 626 tensors and no vision/projector tensors. Gemma 3 vision support is provided by the shared mmproj/mmproj-Gemma-3-12B-IT-BF16.gguf file. Independent exports from the Balanced and Max masters were byte-identical (SHA-256 e92aa2e34524919ad11372fc3de35dba193376eb8474f3f504f817f9177b7caf).

The intermediate F16 GGUF master was used only during conversion and is not published; Q8_0 is the highest-fidelity Director GGUF in this release.

Full Transformers masters

LTX text encoders

The text-encoder files retain the Gemma tokenizer and the tensor layout expected by the stock ComfyUI Gemma loader. Quantization is performed from the exported Heretic-MOE master, not from another quantized artifact.

Hardware table

Measured values will replace the placeholders below after the final files are loaded in real workflows.

Validation matrix

Repository layout

GGUF/
  balanced/
  max/
mmproj/
Text_Encoder/
  balanced/
  max/
Transformers/
  balanced/
  max/
research/
  manifests/
  validation/

Reproducibility

The release pipeline is maintained in the Heretic-MOE repository:

• research/scripts/build_gemma3_ltx_te.py
• research/scripts/quantize_gemma3_ltx_te.py
• research/scripts/run_remote_gemma3_release_pipeline.sh

Every published model file will be accompanied by its source trial, conversion configuration, SHA-256 checksum, and validation status.

Current status

Balanced trial 555 and Max trial 752 are being exported independently on two GPUs. Quantized files remain marked as building until their local checksum, remote LFS checksum, and format-specific load validation pass.