simpledirect/Vinci-Bozza-1.0

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 8, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Vinci Bozza 1.0 is a 9-billion-parameter open-weight model developed by SimpleDirect, fine-tuned from Qwen 3.5-9B. This model focuses on disposition tuning rather than raw capability, aiming for increased honesty, directness, and refusal to fabricate information. It features a 262,144 token context length and is optimized for general conversation, drafting, and everyday reasoning, while maintaining strong safety metrics.

Loading preview...

Vinci Bozza 1.0: A Disposition-Tuned 9B Model

Vinci Bozza 1.0, developed by SimpleDirect, is a 9-billion-parameter model fine-tuned from Qwen 3.5-9B. Unlike models focused solely on raw intelligence, Bozza prioritizes honesty, directness, and anti-fabrication, achieved through specific character training rather than a capability retrain. It features a substantial 262,144 token context window.

Key Characteristics & Performance

  • Enhanced Safety: Achieves 0.0% Adversarial Attack Success Rate (ASR) on HarmBench and JailbreakBench, with 100% refusal rate on jailbreaks.
  • Anti-Fabrication: Improved citation integrity (78.4% on CBLRE) and trained to decline fabricating answers it doesn't have.
  • General Knowledge: MMLU scores show a slight improvement to 71.6%, indicating general ability is preserved.
  • Tool Use: Demonstrates strong tool selection (93.8% on BFCL live relevance) but shows regression in tool abstention and multi-turn task holding.

Intended Use Cases

  • Conversation & Drafting: Ideal for general chat, content drafting, and everyday reasoning tasks.
  • Honest Responses: Designed to provide direct answers and acknowledge its limitations, making it suitable for applications where factual integrity is paramount.

Limitations

  • Strict Instruction Following: Regressed in literal instruction following (IFEval fell to 73.0%), making it less suitable for rigid-format generation like exact JSON.
  • Agentic Loops: Not recommended for autonomous agentic loops due to regressions in instruction following and tool abstention.
  • Specialized Tasks: Not optimized for frontier coding, advanced mathematical reasoning, or francophone legal work.

Finetunes
2 models