This is a decensored version of Darkhn/Mistral-Medium-3.5-128B-BF16-Text-Only, made using Heretic v1.3.0
[!TIP] This model is reproducible!
See the README in the
reproducedirectory for more information.
Abliteration parameters
| Parameter | Value |
|---|---|
| direction_index | 40.13 |
| attn.o_proj.max_weight | 1.47 |
| attn.o_proj.max_weight_position | 52.50 |
| attn.o_proj.min_weight | 1.39 |
| attn.o_proj.min_weight_distance | 49.92 |
| mlp.down_proj.max_weight | 1.21 |
| mlp.down_proj.max_weight_position | 54.38 |
| mlp.down_proj.min_weight | 0.76 |
| mlp.down_proj.min_weight_distance | 50.68 |
Performance
| Metric | This model | Original model (Darkhn/Mistral-Medium-3.5-128B-BF16-Text-Only) |
|---|---|---|
| KL divergence | 0.0329 | 0 (by definition) |
| Refusals | 8/100 | 93/100 |
Multimodal reconstruction
This release preserves XORTRON's modified language-model tensors while restoring
the untouched Pixtral vision tower and multimodal projector from
mistralai/Mistral-Medium-3.5-128B.
The finished checkpoint uses the native
Mistral3ForConditionalGeneration architecture. All floating tensors are stored
as full FP16 Safetensors, and the repository includes the official processor,
tokenizer, image-token configuration, and multimodal chat template.
The imported vision and projector components were not modified by Heretic. The language model remains the XORTRON derivative.
Featherless compatibility
This repository contains the complete multimodal model in full FP16
Safetensors format using the native Mistral3ForConditionalGeneration
architecture. It includes XORTRON's modified language model together with the
official Pixtral vision tower, multimodal projector, processor configuration,
image token configuration, tokenizer, and multimodal chat template.