migtissera/Tess-4-27B

Overview

migtissera/Tess-4-27B is a 27 billion parameter multimodal model developed by Migel Tissera, based on Qwen/Qwen3.6-27B. It represents the first Tess release in two years and introduces a novel reasoning approach. The model was post-trained using a unique blend of 64K-token long-context agentic traces derived from real engineering work with Fable-5, rather than synthetic data. Its reasoning style was approximated from Fable-5 by a multi-model teacher ensemble comprising Opus-4.8, GPT-5.5, and GLM-5.2, distilled into a coherent voice. This training methodology aims to produce a model that thinks prospectively, verifying and weighing alternatives before acting, rather than merely narrating an answer.

Key Differentiators

• Weight-scaled reasoning: Tess-4-27B allocates deliberation proportionally to problem difficulty, focusing on complex steps like planning, debugging, and synthesis.
• Agentic by design: It supports native, parallel tool use and multi-step problem-solving, capable of building mental models from codebases.
• Long-context capability: Trained on 64K-token long-context agentic traces, enabling it to maintain context over large working sets.
• Multimodal: Inherits Qwen3.6's vision tower, allowing for both text and image inputs.
• Honest and evidence-based: Designed to provide grounded pushback and analysis rather than sycophantic responses.

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 coherence.
• Technical & product judgment: Delivering honest, structured analysis and evidence-based critiques.