ImposterOnline/Pyrex-8B-Instruct-Uncensored

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 1, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ImposterOnline/Pyrex-8B-Instruct-Uncensored is a 7.6 billion parameter instruction-tuned model developed by ImposterOnline, specifically designed for coding tasks. It features a coding-first approach, trained on high-quality code instructions, and supports a 32k context window. This model is uncensored, providing direct answers without refusals, and is optimized for agentic and tool-use workflows, making it suitable for developer work like code generation, bug fixing, and technical explanations.

Loading preview...

Pyrex 8B Instruct Uncensored: A Coding-First, Uncensored LLM

Pyrex 8B Instruct Uncensored, developed by ImposterOnline, is a 7.6 billion parameter instruction-tuned model built with a unique data recipe and QLoRA fine-tuning. It is distinguished by its coding-first design, making it highly effective for developer-centric tasks. The model is also uncensored, providing direct and honest responses without built-in safety-alignment softening.

Key Capabilities

  • Exceptional Code Generation: Trained predominantly on high-quality code instructions (OpenCoder, CodeAlpaca, evol-codealpaca), it produces clean and correct code.
  • Uncensored Responses: Delivers direct answers without refusals, offering unfiltered information.
  • Agentic & Tool-Use Ready: Includes function-calling data in its training, making it well-suited for OpenAI-compatible APIs and agent workflows.
  • Portable: Available in full-precision safetensors and GGUF quants for various platforms like Ollama, llama.cpp, and LM Studio.
  • Strong Benchmarking: Achieves 52.4% pass@1 on HumanEval, a significant improvement of +23.1 points over its base model on unseen problems.

Good For

  • Software Development: Writing code, fixing bugs, and explaining complex technical problems.
  • AI Agents: Powering agent loops and tool-use applications requiring function calling capabilities.
  • Unfiltered Information Retrieval: Use cases where direct, uncensored answers are preferred.
  • Local Deployment: Running efficiently on various hardware configurations via GGUF quantizations.

Finetunes
2 models