Oriel — AI & technology worth noticing, every day.

xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic

AI model · https://huggingface.co/xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic

First seen by Oriel on .

AI summary

A Heretic v2.0.0.dev0-decensored Qwen3.5-9B reasoning model, distilled from DeepSeek-V4-Pro, reporting strong mathematics, STEM, efficiency, and output-format performance.

  • Training used approximately 250,000 mathematics and STEM samples, followed by light GSPO reinforcement learning, with no coding-specific supervision.
  • The model achieved 94.50% on GSM8K and averaged 90.53% across sampled MMLU-Pro Math, Physics, and Chemistry questions.
  • It produced 316.8 correct answers per million reasoning tokens, reported as 56.4% more than Qwen3.5-9B.
  • On MMLU-Pro, 97.2% of responses followed the requested single-letter format, compared with 37.3% for official Qwen3.5-9B.
  • The release is experimental; evaluations used sampled subsets and different inference builds, while coding gains remain preliminary and high-stakes outputs require verification.

#reasoning models #model distillation #qwen #llm benchmarks

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 9,377, latest observation ; +9,243 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 29,567, latest observation ; +28,795 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsxtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretichttps://huggingface.co/xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic Read discussion