Public mirror of the BF16 Heretic checkpoint originally released as coder3101/gemma-4-26B-A4B-it-heretic at commit e6d09d5c434a5e91326ce9d779786a91e7aaae35.

Post-process metadata

The checkpoint was produced from google/gemma-4-26B-A4B-it with Heretic 1.2.0 using Arbitrary-Rank Ablation with row-norm preservation.

Parameter	Value
`start_layer_index`	10
`end_layer_index`	30
`preserve_good_behavior_weight`	0.5480
`steer_bad_behavior_weight`	0.0009
`overcorrect_relative_weight`	0.5868
`neighbor_count`	14

Reported by the original post-process run:

Metric	Heretic	Google source
KL divergence	0.0499	0
Refusals	11/100	100/100

2026-07-16 metadata refresh

Google's current source revision is 01e5b3ee840d3a9e0b0b493c593e85398a30ef75. Its two BF16 weight shards have the same LFS SHA256 values as the initial release revision 47b6801b24d15ff9bcd8c96dfaea0be9ed3a0301:

1127684971bbca40465435a5cad69d67ad603bf5e61c6dfd5561fae4a3bcfdb3
aab47033e1e8a492ef8e581efae1cf36478d0433567e7729b3c1728bc8970db7

Because the source weights did not change, rerunning Heretic would transform the same tensors. This mirror keeps the already verified Heretic weights and refreshes the chat template, tokenizer metadata, and processor metadata from Google's current revision. The refresh includes fixes for tool-call turns, thinking history, null arguments, and multimodal tool responses.

Loading

python

from transformers import AutoModelForMultimodalLM, AutoProcessor

model_id = "WaveCut/gemma-4-26B-A4B-it-heretic"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForMultimodalLM.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)

This is a third-party post-processed checkpoint. It is not affiliated with or endorsed by Google. Review the Apache 2.0 license and Gemma license link above before redistribution or deployment.