migtissera/Tess-4-27B

Overview

migtissera/Tess-4-27B is a 27 billion parameter model developed by Migel Tissera, based on Qwen/Qwen3.6-27B. It is distinguished by its unique post-training on 64K-token long-context agentic traces, which are derived from real engineering work and a multi-model teacher ensemble (Opus-4.8, GPT-5.5, GLM-5.2). This training methodology enables Tess-4-27B to deliberate more intensely on complex problems while being efficient on routine tasks, mimicking the thought process of a senior engineer.

Key Capabilities

• Weight-scaled reasoning: Adapts its deliberation depth based on problem difficulty, focusing on planning, debugging, and synthesis.
• Agentic by design: Supports native, parallel tool use and disciplined multi-step problem solving, including codebase analysis.
• Long-context handling: Trained on 64K-token contexts, allowing it to maintain coherence over large documents and codebases.
• Multimodal: Inherits Qwen3.6's vision capabilities, processing both text and image inputs.
• Honest and evidence-based: Provides grounded pushback rather than sycophantic responses.
• Prospective reasoning: Reasons by predicting, verifying, and weighing alternatives before acting.

Benchmarks and Performance

Tess-4-27B currently holds a best-in-class score for BenchLocal, achieving 81% (122/150) in community-performed benchmarks, outperforming other 27B and 35B 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 critiques.