CJRouter/qwen3.8-27b-clef-decision

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

CJRouter/qwen3.8-27b-clef-decision is a 27 billion parameter decision model based on Qwen3.8-27B, incorporating a Cloudflare Clef-style joint_head architecture. It is specifically fine-tuned to produce structured, calibrated-probability decisions for complex business scenarios. This model excels at tasks requiring structured judgment, such as agent trace observability, customer service ticketing, and security incident response, offering fine-grained per-question accuracy and calibrated confidence.

Loading preview...

Qwen3.8-27B-Clef Decision Model Overview

This model, developed by CJRouter, is a specialized 27 billion parameter decision model built upon the Qwen3.8-27B base. It integrates a Cloudflare Clef-style joint_head architecture, enabling it to generate structured, calibrated-probability decisions for intricate business contexts. The model was efficiently trained in approximately 67 minutes on a single NVIDIA H20 using LoRA and joint decision head fine-tuning.

Key Capabilities

  • Structured Decision-Making: Outputs choice probability distributions and confidence for defined questions based on state descriptions.
  • Calibrated Confidence: Achieves an Expected Calibration Error (ECE) of 0.0291 on an independent test set, ensuring that stated confidence levels accurately reflect correctness.
  • Fine-Grained Per-Question Accuracy: Evaluated with per-question accuracy, reaching 76-87% across various business domains, allowing for individual actionability and confidence thresholds.
  • Rapid Training: Reproducible training on commodity hardware in about an hour, demonstrating efficiency and accessibility.
  • Private and Sovereign: Licensed under Apache-2.0, designed for on-premise deployment without vendor lock-in or data egress.

Good For

  • Agent Trace Observability: Judging autonomous agent run traces for continuation, human intervention needs, task outcomes, and risk levels.
  • Customer Service Ticketing: Automating determination of ticket urgency, owning team, and customer sentiment.
  • Invoice Processing: Making structured judgments on invoice-related tasks.
  • Security Incident Response: Assessing incident risk levels and handling priorities with calibrated confidence.
  • Applications requiring defensible, automated risk-threshold decisions.

Finetunes
1 models