KRLabsOrg/squeez-2b

VISIONConcurrent Unit Cost:1Model Size:2.3BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 16, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

KRLabsOrg/squeez-2b is a 2.3 billion parameter model fine-tuned from Qwen 3.5 2B, specifically designed for task-conditioned tool-output pruning in coding agents. It excels at extracting the smallest verbatim evidence block from verbose tool outputs, removing 92% of input tokens while retaining 0.86 recall. This model significantly improves the efficiency of coding agents by providing focused context, outperforming larger zero-shot models like Qwen 3.5 35B A3B in recall for software engineering tasks.

Loading preview...

What is Squeez-2B?

Squeez-2B is a 2.3 billion parameter language model developed by KRLabsOrg, fine-tuned from Qwen 3.5 2B. Its primary function is task-conditioned tool-output pruning for coding agents. This means it takes a focused query and a raw tool observation (like logs or command outputs) and extracts only the most relevant lines, significantly reducing the input size for subsequent agent processing.

Key Capabilities & Features

  • Efficient Context Pruning: Reduces tool output by 92% while maintaining 0.86 recall, providing agents with highly focused information.
  • Verbatim Extraction: Returns exact lines from the original output without rewriting or summarization, preserving original context.
  • Superior Performance: Outperforms zero-shot Qwen 3.5 35B A3B by +11 recall points in tool-output extraction.
  • Broad Tool Type Support: Trained on 27 diverse tool types from real SWE-bench workflows and synthetic multi-ecosystem outputs.
  • Flexible Deployment: Can be used as a CLI pipe, Python library, or vLLM server.

Why use Squeez-2B?

Squeez-2B is ideal for enhancing coding agents by making their processing of tool outputs more efficient and accurate. It helps agents focus on critical information, reducing noise and improving decision-making. Its small size (2.3B parameters) combined with its specialized performance makes it a highly effective and resource-friendly solution for specific software engineering tasks, particularly when dealing with verbose logs or command outputs. It is not designed for general-purpose summarization but for precise, task-specific evidence extraction.