barozp/Qwen3.8-27B-Opus-Distill-v2

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

barozp/Qwen3.8-27B-Opus-Distill-v2 is a 27 billion parameter language model fine-tuned from Qwen/Qwen3.8-27B. It is specifically distilled from verified Claude Opus reasoning traces to enhance reasoning quality while fixing a critical bug present in its predecessor. This model excels in complex reasoning tasks, as evidenced by significant gains on benchmarks like GPQA-Diamond, and is suitable for applications requiring robust logical processing and adherence to stacked output constraints.

Loading preview...

Model Overview

barozp/Qwen3.8-27B-Opus-Distill-v2 is a 27 billion parameter model based on Qwen/Qwen3.8-27B, fine-tuned using LoRA. This version specifically addresses and resolves a critical reasoning-loop bug found in its predecessor (v1) where stacked output constraints could lead to token exhaustion without visible output. The fix involved meticulously reconstructing the training data, replacing fabricated reasoning traces with 100% verified-genuine Claude Opus traces from sources like lordx64 and Roman1111111.

Key Enhancements and Features

  • Bug Fix: Eliminates the non-converging self-verification loop when faced with stacked output constraints (e.g., "no prose" + "no markdown"), ensuring clean and concise output.
  • Reasoning Quality: Maintains significant gains in reasoning capabilities, with a notable +0.237 improvement on GPQA-Diamond and +0.034 on ARC-Challenge compared to the base model.
  • Data Integrity: Trained on a refined dataset of 11,716 verified-genuine Opus traces, ensuring higher quality distillation.
  • Base Model Capabilities: Inherits native vision-language and Multi-Turn Prediction (MTP) capabilities from the Qwen3.8-27B base model, though these were not explicitly trained in this distillation.

Ideal Use Cases

  • Applications requiring strong logical reasoning and problem-solving.
  • Scenarios where strict adherence to multiple output formatting constraints is crucial.
  • Tasks benefiting from Claude Opus-like reasoning patterns without the associated cost or API dependency.
  • Developers seeking a robust 27B model with enhanced reasoning and reliability for whom the v1 bug was a blocker.

Finetunes
7 models