DavidAU/ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking
DavidAU/ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking is a 21 billion parameter Mixture-of-Experts (MoE) model, fine-tuned from Baidu's ERNIE-21B-A3B using a Claude 4.5 High Opus reasoning dataset. This model features a 128k context window and is optimized for deep reasoning, creative tasks, and general usage, offering detailed and compact reasoning capabilities. It demonstrates improved benchmark performance over its base model in areas like ARC Challenge and HellaSwag.
Loading preview...
Model Overview
DavidAU/ERNIE-21B-A3B-Claude-4.5-High-OPUS-Thinking is a 21 billion parameter Mixture-of-Experts (MoE) model, specifically a fine-tune of Baidu's ERNIE-21B-A3B. It was trained using a high-quality reasoning dataset derived from Claude 4.5 High Opus, leveraging Unsloth for efficient fine-tuning. A key characteristic of this model is its "uncensored" nature and focus on deep, detailed reasoning.
Key Capabilities & Features
- Enhanced Reasoning: The model is designed for "full deep thinking," providing compact yet very detailed reasoning outputs.
- Creative & General Usage: Excels in creative tasks such as brainstorming and prose generation, alongside strong general-purpose performance.
- Extended Context Window: Supports a 128k token context length.
- Temperature Stability: Reasoning capabilities are stable across a wide temperature range (0.1 to 2.5).
- Improved Benchmarks: Demonstrates notable performance gains over the regular ERNIE-21B-A3B model, with scores like 0.392 on ARC Challenge (vs 0.331) and 0.686 on HellaSwag (vs 0.663).
Optimal Usage & Settings
To achieve the best performance, specific settings are recommended:
- Repetition Penalty: For creative tasks, a repetition penalty of 1.01 to 1.1 is suggested. For general work, 1.05 or 1.1, or even off (1).
- Smoothing Factor: Setting a "Smoothing_factor" to 1.5 in interfaces like KoboldCpp, oobabooga/text-generation-webui, or Silly Tavern is highly recommended for smoother operation, especially for chat and roleplay. This can often negate the need for increased repetition penalty.
- Quadratic Sampling: If supported by the interface, utilizing "Quadratic Sampling" (often referred to as "smoothing") is beneficial.
Lower quantization levels may occasionally lead to looping behavior. For comprehensive guidance on advanced settings and maximizing performance across various use cases, users are directed to the Maximizing Model Performance guide.