Michael-Kozu/Kuiper-R1

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 10, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Michael-Kozu's Kuiper-R1 is a 9.41 billion parameter Qwen3.5-based conditional generation model designed for reasoning-trace inversion. It reconstructs full, readable reasoning from a problem, its final answer, and compressed reasoning bubbles, preserving the final answer byte-for-byte. This model is optimized for reasoning-dataset augmentation, interpretability review, and research into reasoning compression and expansion, with a notable 262,144 token context length.

Loading preview...

Overview

Kuiper-R1 by Michael-Kozu is a 9.41 billion parameter model built on the Qwen3.5 architecture, specializing in reasoning-trace inversion. Its core function is to reconstruct detailed, human-readable reasoning from a given problem, its final answer, and a set of compressed "reasoning bubbles." The model ensures the final answer is preserved exactly, while the reconstructed trace is explicitly marked as synthetic.

Key Capabilities

  • Reasoning Reconstruction: Generates full reasoning traces from partial inputs (problem, answer, reasoning bubbles).
  • Answer Preservation: Guarantees byte-for-byte preservation of the final answer.
  • High Context Length: Features a substantial 262,144 token context window.
  • Structured Output: Adheres to a three-tag contract: <synthetic_trace>, <verification_note>, and <final>.

Training and Performance

The model was trained using QLoRA SFT and ORPO methods on 3,000 SFT examples and 87 preference pairs from the OpenThoughts3-1.2M dataset, focusing on science, math, and code. Benchmarks show an 82% exact answer preservation rate and 88% format validity, with an overall score of 50.4% on a held-out dataset.

Good For

  • Reasoning Dataset Augmentation: Expanding compressed reasoning into full traces.
  • Interpretability Research: Studying how reasoning can be compressed and expanded.
  • Research into Reasoning Compression: Investigating methods for efficient reasoning representation.

Limitations

  • Trace quality is dependent on the informativeness of input bubbles.
  • The verification note is an internal consistency statement, not an external validation.
  • While improving answer preservation, it may slightly trail base models in raw format validity and semantic consistency.

Finetunes
1 models