schneewolflabs/B1.1-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 6, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

schneewolflabs/B1.1-9B is a 9 billion parameter language model developed by Schneewolf Labs, featuring a 32768 token context length. It is an iteration of the B1-9B model, with its ORPO adapter merged at half strength to balance persona and tool-use capabilities. This model is optimized for improved agentic behavior, particularly in tool-use scenarios, while mitigating persona drift and maintaining strong code implementation skills.

Loading preview...

Schneewolf Labs B1.1-9B Overview

Schneewolf Labs B1.1-9B is a 9 billion parameter language model, an iterative refinement of the B1-9B model. It incorporates an ORPO adapter from the Vernunft-Stimme dataset, merged at half strength (0.5). This specific merge scale was chosen after observing that a full merge (1.0) in B1-9B, while fixing an 'empty-answer-after-' bug, caused significant persona drift towards the base model. B1.1-9B aims to retain the critical fix for agentic thinking while recovering much of the original prose quality and reducing unwanted censorship, offering a more balanced performance.

Key Capabilities & Improvements

  • Enhanced Agentic Thinking: Significantly improves the ability to answer after a </think> token, achieving 5/8 success rate in native tool tests, compared to 0/8 in B0-9B.
  • Balanced Persona: Recovers prose distance (0.679 vs 1.881 for B1-9B) and reduces censorship (27/29 vs 25/29 for B1-9B), mitigating the persona drift seen in B1-9B.
  • Strong Code Implementation: Maintains high performance on code tasks, scoring 9/10 on the buchbinder ladder for implementing functions.
  • Tool Use: While showing some minor tool-choice drift inherited from B1, it largely retains the egirl 47-case tool bench performance.
  • Vision Capabilities: The vision tower remains intact, with mmproj included, ensuring multimodal functionality.

Good For

  • Agentic Applications: Ideal for use cases requiring models to process thoughts and utilize tools effectively, especially when integrated with native tools field and without legacy /think prefixes.
  • Code Generation & Implementation: Suitable for tasks involving writing and verifying code functions.
  • Balanced Generative Tasks: When a model is needed that combines improved agentic behavior with a more neutral persona and good prose quality.
  • Multimodal Applications: For scenarios leveraging its intact vision capabilities.

Finetunes
2 models