violetxi/qwen35-9b-harvey-v4-notes-conditioned-10m

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-10m is a 9 billion parameter language model based on the Qwen3.5-9B architecture, developed by violetxi. It is specifically fine-tuned using a unique 10 million token dataset comprising 70% note labels and 30% note-conditioned trajectory assistant labels. This model is optimized for agentic tasks requiring historical thinking and complex reasoning, as evidenced by its evaluation on multi-turn historical agent generations. Its training methodology focuses on causal next-token prediction for notes and trajectory loss for assistant labels, making it suitable for applications demanding structured reasoning and problem-solving.

Loading preview...

Model Overview

The violetxi/qwen35-9b-harvey-v4-notes-conditioned-10m is a 9 billion parameter language model built upon the Qwen3.5-9B base. This specific checkpoint, checkpoint-612, is the result of a specialized fine-tuning experiment involving 10 million supervised tokens over two epochs.

Key Training Details

  • Base Model: Qwen/Qwen3.5-9B.
  • Dataset: Fine-tuned on 10,001,435 supervised tokens per epoch, consisting of a 70% note label and 30% note-conditioned trajectory assistant label mixture.
  • Methodology: Utilizes causal next-token prediction for notes and applies trajectory loss to assistant labels, with no KL regularization.
  • Configuration: Trained with a packed sequence length of 16,384, a global batch size of 4, and a learning rate of 5e-06.

Evaluation and Performance

The model's performance was evaluated on historical thinking-enabled agent generations using a gpt-5.6-sol regrading system. It measures the fraction of attempts passing all criteria across 250 tasks with 4 samples each.

  • 5-Turn Budget: Achieved a 3.30% all-criteria-pass rate.
  • 20-Turn Budget: Achieved a 6.20% all-criteria-pass rate.

These evaluations used a historical 5/20-turn glob/grep/read protocol, indicating its capability in agentic reasoning tasks. Further closed-book recall probes were conducted on 7,933 items, with detailed metrics available in the evaluation dataset.

Ideal Use Cases

This model is particularly well-suited for:

  • Agentic AI Systems: Developing agents that require structured reasoning and historical context.
  • Complex Problem Solving: Applications demanding multi-step thought processes and adherence to specific criteria.
  • Research in Agent Behavior: Exploring and analyzing agent performance in constrained environments.