richardyoung/Mythos-nano-heretic
The richardyoung/Mythos-nano-heretic is a 3.1 billion parameter language model, a decensored version of squ11z1/Mythos-nano created using Heretic v1.4.0. This model is specifically engineered to reduce refusals and safety guardrails, making it suitable for use cases where a less restricted output is desired. It demonstrates strong performance in competitive programming and complex mathematical reasoning, often rivaling much larger models in these specific domains.
Loading preview...
Mythos-nano-heretic: A Decensored 3.1B Parameter Model
This model, developed by richardyoung, is a decensored variant of the original squ11z1/Mythos-nano model, created using Heretic v1.4.0. Its primary distinction lies in the significant reduction of refusal behaviors, with only 13 refusals out of 100 compared to the original's 79/100. This makes it suitable for applications requiring less constrained outputs, though users are advised to exercise caution due to reduced safety guardrails.
Key Capabilities & Performance
Despite its compact 3.1 billion parameters, Mythos-nano-heretic exhibits remarkable performance, particularly in:
- Competitive Mathematics: Achieves high scores on benchmarks like AIME, HMMT, and BruMO, often performing comparably to models with hundreds of billions or even trillions of parameters.
- Competitive Programming: Demonstrates a 96.1% pass-rate on LeetCode contests (Python), placing it near top-tier models like GPT-5.3-Codex and Gemini 3.1 Pro.
Important Considerations
- Decensored Nature: The model's refusal direction has been removed, meaning it will not decline requests that a safety-tuned model typically would. Users are solely responsible for its outputs and legal compliance.
- Not for Tool-Calling: It is explicitly not recommended for tasks involving function calling, API orchestration, or autonomous coding agents, as it was not trained on such data.
- Reproducibility: The model's creation process is reproducible, with details available in the
reproduce/README.md.
Usage
Recommended sampling parameters include a temperature of 0.6–1.0 and up to 40960 output tokens for challenging problems. GGUF formats (mythos-nano-f16.gguf and mythos-nano-Q4_K_M.gguf) are provided for llama.cpp and Ollama.