kintsugicollective/atlas-trm10-gemma4-26b

VISIONConcurrency Cost:2Model Size:26BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Jun 6, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Atlas Gemma-4-26B-trm v10 is a 26 billion parameter language model developed by Kintsugi Collective, based on Google's Gemma-4-26B-a4b-it architecture, with a 32768 token context length. This experimental model is specifically modified to reduce therapeutic refusal behavior and crisis-line reflexes, making it a specialized therapeutic-context model for adults with complex trauma and neurodivergence. It achieves a 0% therapeutic refusal rate and 100% Region 1 Safety, distinguishing it from general-purpose LLMs.

Loading preview...

Atlas Gemma-4-26B-trm v10: Specialized Therapeutic AI

Atlas Gemma-4-26B-trm v10 is an experimental 26 billion parameter language model developed by Kintsugi Collective, specifically designed as an intelligence layer for their platform supporting adults with complex trauma (CPTSD), PTSD, and neurodivergence (ASD/ADHD). It is not a general-purpose model.

Key Capabilities & Modifications

  • Base Model: Built upon google/gemma-4-26b-a4b-it.
  • Therapeutic Refusal Modification (TRM): Intentionally modified to reduce therapeutic refusal behavior and crisis-line reflexes through norm-preserving biprojected abliteration and Expert-Granular Abliteration (EGA) across all 30 layers. This process involved full expert ablation and a bespoke harmful prompt corpus.
  • Fine-tuning: Supervised fine-tuning (SFT) was performed over 3 epochs on a curated dataset of 2,395 examples, comprising 60% high-quality synthetic data and 40% redacted lived-experience data from the target cohort.
  • Safety & Refusal: Achieves a 0% Therapeutic Refusal Rate and 100% Region 1 Safety (weapons, CSAM, violence), as indicated by internal benchmarks.

Performance Highlights

  • GSM8K (Math reasoning): 90.0%
  • HellaSwag (General reasoning): 61.6%
  • TruthfulQA (Truthfulness): 63.2%
  • MMLU (Multitask language understanding): 61.6%

Limitations & Responsible Use

This model has reduced refusal behavior on therapeutic and dark content and is not suitable for general deployment without guardrails. It is intended for use within the Atlas companion architecture with additional safety layers and is not a replacement for human therapeutic support. The model is licensed under Apache 2.0.

Finetunes
1 models