hrktos-37/Hermes-4-70B-heretic

hrktos-37/Hermes-4-70B-heretic: Decensored Reasoning Model

This model is a 70 billion parameter, decensored variant of NousResearch's Hermes-4-70B, built upon the Llama-3.1-70B architecture. It was created using Heretic v1.1.0 to significantly reduce refusal rates, demonstrating 26/100 refusals compared to the original model's 47/100 on the RefusalBench benchmark. The model maintains a 32768 token context length and is designed for enhanced helpfulness and user alignment.

Key Capabilities

• Advanced Reasoning: Features a "hybrid reasoning mode" with explicit <think>…</think> segments for deliberation, improving performance in math, code, STEM, and logic.
• Improved Steerability: Offers extreme improvements in steerability and reduced refusal rates, making it highly adaptable to user values and preferences.
• Structured Outputs: Trained for robust schema adherence, capable of producing valid JSON and repairing malformed objects.
• Enhanced Training: Benefits from a post-training corpus of ~5M samples / ~60B tokens, emphasizing verified reasoning traces.
• Function Calling & Tool Use: Supports tool calls within a single assistant turn, integrating seamlessly with reasoning mode for improved accuracy.

Good for

• Applications requiring a highly steerable and less censored large language model.
• Tasks demanding strong reasoning capabilities, including complex problem-solving in math, code, and STEM.
• Generating structured outputs like JSON, ensuring format-faithful responses.
• Creative writing and subjective response generation where expressive freedom is desired.
• Developers looking for a model with explicit internal reasoning processes for debugging or transparency.