violetxi/qwen35-9b-harvey-v4-notes-conditioned-5m

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The violetxi/qwen35-9b-harvey-v4-notes-conditioned-5m is a 9 billion parameter Qwen3.5-based language model, fine-tuned with a 32768 token context length. It was trained using a unique 70% note labels and 30% note-conditioned trajectory assistant labels mixture, totaling over 5 million supervised tokens. This model is specifically designed for agentic tasks requiring historical thinking, demonstrating its capabilities in complex multi-turn evaluations.

Loading preview...

Overview

This model, qwen35-9b-harvey-v4-notes-conditioned-5m, is a 9 billion parameter variant of the Qwen3.5 architecture, fine-tuned for specialized agentic reasoning. It leverages a unique training methodology involving a 70% note labels and 30% note-conditioned trajectory assistant labels mixture, totaling over 5 million supervised tokens across two epochs. The training utilized a packed sequence length of 16,384 tokens and a learning rate of 5e-06.

Key Capabilities

  • Specialized Fine-tuning: Conditioned on "notes" and "note-conditioned trajectories" to enhance specific reasoning patterns.
  • Agentic Task Performance: Evaluated on historical thinking-enabled agent generations, achieving a 3.40% all-criteria-pass rate on 20-turn tasks using a glob/grep/read protocol.
  • Qwen3.5 Base: Built upon the robust Qwen3.5-9B foundation, inheriting its general language understanding capabilities.

Good For

  • Agent Development: Ideal for researchers and developers building agents that require nuanced reasoning and historical context processing.
  • Specialized Reasoning Tasks: Particularly suited for applications where models need to process and generate responses based on internal 'notes' or structured thought processes.
  • Experimental AI Systems: Useful for exploring the impact of note-conditioned training on model behavior and performance in complex, multi-turn interactions.