Babsie/DavidAU_L3.3_8B_OpusR_HUC

TEXT GENERATIONPricing:Input $0.37 / Cached $0.074 / Output $0.38Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Mar 5, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Babsie/DavidAU_L3.3_8B_OpusR_HUC is an 8 billion parameter Llama 3.3-based instruction-tuned model developed by DavidAU, featuring a 128k context length. This model is 'Heretic'd' (de-censored) and further trained with a Claude 4.5-Opus High Reasoning dataset, creating an uncensored instruct/thinking hybrid. It is specifically designed for neuroatypical accessibility and excels at complex reasoning tasks, automatically activating a 'thinking' process for detailed responses.

Loading preview...

Model Overview

Babsie/DavidAU_L3.3_8B_OpusR_HUC is an 8 billion parameter model based on a "Llama3.3-8B" source, adjusted to a 128k context length. Developed by DavidAU, this model has undergone significant modifications to enhance its reasoning and instruction-following capabilities while being fully uncensored.

Key Capabilities

  • Uncensored Output: The model has been "Heretic'd" to remove original censorship and trained to reduce refusals (14/100 refusal rate, KL divergence 0.1498).
  • High Reasoning: Fine-tuned with a Claude 4.5-Opus High Reasoning dataset, it functions as an instruct/thinking hybrid.
  • Automatic Thinking Activation: Certain phrases or words in prompts (e.g., "Think Deeply:") automatically activate a detailed internal "thinking" process, leading to more elaborate and structured responses.
  • Extended Context: Supports a 128k context window, with a suggested minimum context window of 4k and optimal performance at 8k+.

Good For

  • Complex Reasoning Tasks: Ideal for prompts requiring detailed explanations, mathematical examples, or structured creative writing where the "thinking" activation can be leveraged.
  • Unrestricted Content Generation: Suitable for use cases where uncensored or explicit content generation is desired, though it may require explicit directives (e.g., using specific slang or terms) to achieve the desired graphic level.
  • Neuroatypical Accessibility: Developed with a focus on accessibility for neuroatypical users.

Benchmarks

Preliminary benchmarks show scores such as ARC-Challenge: 0.480, ARC-Easy: 0.687, BoolQ: 0.831, HellaSwag: 0.705, OpenBookQA: 0.438, PIQA: 0.780, and Winogrande: 0.646.