schneewolflabs/B2-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 27, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

B2-9B is a 9-billion parameter language model developed by Schneewolf Labs, built upon the B0-9B base. It is specifically fine-tuned for agentic tasks within the egirl harness, emphasizing careful, truth-telling, and ask-first behavior to prevent data loss. With a 32768-token context length, this model excels at reliable tool use and significantly reduces destructive actions, making it suitable for sensitive operational environments.

Loading preview...

Schneewolf Labs B2-9B Overview

B2-9B is a 9-billion parameter model from Schneewolf Labs, designed to operate as a careful and reliable agent within the egirl harness. It builds upon the B0-9B base model, undergoing Supervised Fine-Tuning (SFT) and Orthogonal Reinforcement Learning (ORPO) to instill specific operational behaviors.

Key Capabilities

  • Enhanced Agentic Reliability: Significantly reduces instances of empty answers after tool results and destructive actions, particularly preventing the loss of irreplaceable data (0/24 cases in held-out tests).
  • Careful Tool Use: Trained to ask before performing potentially destructive actions and to verify outcomes, demonstrating honesty in reporting test reruns.
  • Reduced Stance and Plainer Voice: Exhibits a 0% stance rate, indicating fewer opinions, and a plainer prose style, making it a quieter and more objective operator.
  • Improved Safety: Maintains strong safety asymmetry, refusing actual harmful requests.
  • Optimized for egirl: Specifically trained on Qwen3.5's native <function=…> format, which egirl sends, ensuring compatibility and effective tool interaction.

Good For

  • Agentic Workflows: Ideal for applications requiring a highly reliable and cautious agent that interacts with tools and sensitive data.
  • Operational Environments: Suitable for tasks where preventing data loss and ensuring precise execution are critical.
  • Automated Task Execution: Useful for scenarios where an agent needs to think through steps, confirm actions, and provide truthful feedback without expressing personal opinions.

Finetunes
2 models