pb09204048/CRISP-Qwen3-8B-v2

Overview

pb09204048/CRISP-Qwen3-8B-v2 is an 8 billion parameter model based on the Qwen3 architecture, fine-tuned using the CRISP (Compressed Reasoning via Iterative Self-Policy Distillation) method. This specific version utilizes a v2 conciseness teacher, which is designed to be difficulty-aware, preventing over-compression of complex, multi-step problems. The CRISP method teaches the model to generate concise reasoning by distilling its own concise behavior, minimizing per-token reverse KL divergence from a self-conditioned teacher to a student model.

Key Capabilities & Performance

This model excels at producing more concise reasoning outputs while largely preserving accuracy across various benchmarks. Compared to its base Qwen3-8B model, CRISP-Qwen3-8B-v2 demonstrates significant token reduction:

• MATH-500: Achieves 95.7% accuracy with a 31.6% token reduction.
• AIME 2024: Maintains 75.0% accuracy with a 17.1% token reduction.
• AIME 2025: Shows 65.8% accuracy with a 17.5% token reduction.
• GPQA-Diamond: Scores 58.3% accuracy with a 17.2% token reduction.
• MMLU: Achieves 81.2% accuracy with a 22.4% token reduction.

These results indicate that the model can generate more efficient reasoning steps without a substantial drop in performance, particularly in mathematical and general knowledge tasks, under a 30K-token budget.

Ideal Use Cases

• Efficient Reasoning: Applications requiring concise explanations or problem-solving steps to save computational resources or improve readability.
• Mathematical Problem Solving: Tasks involving complex math or logical reasoning where brevity is desired without sacrificing correctness.
• Resource-Constrained Environments: Scenarios where token usage needs to be minimized due to cost or latency considerations.