migtissera/Tess-4-9B
Tess-4-9B is a 9 billion parameter agentic language model developed by Migel Tissera, built upon Qwen/Qwen3.5-9B-Base. It is post-trained on 64K-token long-context agentic traces, enabling weight-scaled reasoning that allocates deliberation proportionally to task difficulty. This model excels at complex problem-solving, agentic coding, and long-context work, offering a practical footprint for deployment.
Loading preview...
Overview
Tess-4-9B is a 9 billion parameter model developed by Migel Tissera, serving as a compact version of Tess-4-27B. It is built on Qwen/Qwen3.5-9B-Base and post-trained on 64K-token long-context agentic traces, which are real engineering workflows rather than synthetic generations. The model's reasoning style is approximated from Fable-5 using a multi-model teacher ensemble (Opus-4.8, GPT-5.5, GLM-5.2), resulting in a model that thinks prospectively, verifying and weighing alternatives before acting.
Key Capabilities
- Weight-scaled reasoning: The model adapts its deliberation effort to the difficulty of the task, focusing on planning, debugging, and synthesis for hard problems while keeping routine steps concise.
- Agentic by design: Supports native, parallel tool use and disciplined multi-step problem-solving, capable of building mental models from codebases.
- Long-context processing: Trained on 64K-token traces, allowing it to maintain context over large working sets.
- Multimodal: Inherits Qwen3.5's vision tower, supporting both text and image inputs.
- Honest and evidence-based: Trained to provide grounded pushback with evidence rather than sycophantic responses.
Performance Highlights
Tess-4-9B significantly improves upon its base model, Qwen3.5-9B-Base, across various benchmarks:
- MMLU: Achieves 79.4%, a 56.2-point increase over the base.
- GSM8K: Scores 88.0% in strict accuracy, a 37-point improvement.
- GPQA Diamond: Reaches 62% with a 24K generation cap, extending to 73% with extended context.
Good For
- Agentic coding: Ideal for exploring unfamiliar repositories, planning changes, and executing multi-step tasks with tools.
- Long-context work: Effective for reasoning over extensive codebases and documents.
- Hard reasoning: Suitable for demanding math, science, debugging, and synthesis tasks that require substantial inference-time thought.
- Technical and product judgment: Provides honest, structured analysis and evidence-based pushback.
- Local and single-GPU experimentation: Its 9B parameter size (approx. 19 GB BF16) makes it substantially easier to deploy and iterate with compared to larger models.