morriszjm/Tacit-9B

Overview

Tacit-9B is a 9-billion parameter language model, based on Qwen/Qwen3.5-9B, uniquely optimized for making typed decisions (e.g., choosing an option, answering yes/no, or selecting from ordered levels) in a single forward pass. Its development utilized AnyJev self-distillation, a process where the base model generated its own decision problems and learned to provide direct answers without human labeling or external datasets.

Key Capabilities

• Efficient Decision Making: Returns a probability distribution over predefined options in a single forward pass, making it highly efficient for classification-like tasks.
• Adaptive Reasoning: Features an optional adaptive mode where low-confidence decisions (based on a configurable tau threshold) are routed to the model's own reasoning process (thinking on), improving accuracy for challenging cases.
• Benchmarked Performance: Demonstrates improved accuracy over its base model, Qwen3.5-9B, on decision-focused benchmarks:
  • JevBench (public set): +1.7 points (0.823 for Tacit-9B)
  • bev-decision (test split): +4.5 points (0.725 for Tacit-9B)
• Flexible Decision Types: Supports "choice", "yes_no", and "score" (ordered levels) decision kinds.
• vLLM Integration: Optimized for high-throughput inference with vLLM, supporting both direct integration and server-client architectures.

Good for

• Automated Classification: Ideal for applications requiring rapid, probabilistic classification or decision-making from a predefined set of options.
• Customer Support Automation: Can be used to quickly categorize customer inquiries (e.g., "track_order", "refund") or determine sentiment.
• Policy Enforcement: Suitable for systems that need to make consistent, rule-based decisions with an option for deeper reasoning on ambiguous cases.
• High-Throughput Inference: Its vLLM compatibility makes it well-suited for production environments demanding fast and scalable decision processing.