Model Overview
DavidAU/Llama-3.3-8B-Instruct-Thinking-Claude-Haiku-4.5-High-Reasoning-1700x is an 8 billion parameter Llama 3.3-based model, distinguished by its extended 128k context window and specialized instruction tuning. The model was fine-tuned using the Claude-Haiku-4.5-high-reasoning-1700x dataset, which imbues it with a unique "thinking" capability. This allows the model to generate internal reasoning processes, often producing short, blunt, and direct thought blocks before formulating its final response.
Key Capabilities
- Enhanced Reasoning: Specifically trained to "think" like Claude-Haiku-4.5, providing structured reasoning for complex prompts.
- Extended Context: Features a 128k context window, suitable for processing and generating longer texts.
- Automatic Thinking Activation: Certain keywords like "explain," "come up with a plan to...", or "think deeply" automatically trigger the model's reasoning process.
- Knowledge Update: The fine-tuning process also updated some of the model's core knowledge and root training.
Good For
- Complex Problem Solving: Ideal for tasks requiring detailed explanations, planning, or deep analytical thought.
- Creative Writing with Structure: Can generate structured creative content, such as detailed story plots, by first outlining a plan.
- Technical Explanations: Excels at breaking down and explaining intricate technical concepts.
Usage Notes
- The model is designed to activate its thinking process automatically for many prompts, but a specific system prompt can be used to force this behavior.
- Suggested settings include a temperature of 0.7, repetition penalty of 1.05, top_p of 0.95, min_p of 0.05, and top_k of 40. A minimum context window of 4k is recommended, with 8k+ preferred.
- For smoother operation, users of KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern may benefit from setting a "Smoothing_factor" to 1.5.