s3nh/Tess-4-27B-abliterated
s3nh/Tess-4-27B-abliterated is a 27 billion parameter decensored version of Migel Tissera's Tess-4-27B model, built upon Qwen/Qwen3.6-27B. This model is specifically designed for agentic, reasoning-intensive tasks, excelling at complex problem-solving by deliberating proportionally to task difficulty. It features a 32K token context length and is optimized for applications requiring deep technical judgment and multi-step agentic workflows, particularly in coding and long-context analysis.
Loading preview...
Overview
s3nh/Tess-4-27B-abliterated is a 27 billion parameter language model, a decensored variant of Migel Tissera's Tess-4-27B, which is based on Qwen/Qwen3.6-27B. This model is distinguished by its "weight-scaled reasoning" approach, meaning it allocates more computational effort to difficult parts of a problem and less to routine steps. It was post-trained on 64K-token long-context agentic traces, derived from real engineering work and a multi-model teacher ensemble (Opus-4.8, GPT-5.5, GLM-5.2) to approximate a senior engineer's reasoning style. The model also inherits Qwen3.6's vision capabilities, supporting both text and image inputs.
Key Capabilities
- Weight-scaled Reasoning: Deliberates proportionally to the difficulty of the task, focusing on planning, debugging, and synthesis.
- Agentic Design: Supports native, parallel tool use and disciplined multi-step problem solving, capable of building mental models from codebases.
- Long-Context Handling: Trained on 64K-token long-context agentic traces, maintaining coherence over large working sets.
- Multimodal: Inherits vision capabilities from Qwen3.6, allowing for image input alongside text.
- Honest & Evidence-Based: Trained to provide grounded pushback and structured analysis rather than sycophantic responses.
- Reproducible Decensoring: Created using Heretic v1.4.0, with specified abliteration parameters.
Performance & Benchmarks
This abliterated version shows a KL divergence of 0.0006 compared to the original Tess-4-27B, with refusals at 81/100 compared to the original's 90/100. The base Tess-4-27B model has demonstrated strong performance in community benchmarks like BenchLocal, achieving 81% (122/150) in its Q8 quantized form, ranking highly among similar models.
Good For
- Agentic Coding: Exploring unfamiliar repositories, planning changes, and executing multi-step tasks with tools.
- Long-Context Work: Reasoning over extensive codebases and documents without losing context.
- Technical & Product Judgment: Providing honest, structured analysis and evidence-based pushback.
- Multimodal Applications: Integrating image and text inputs for comprehensive understanding.