void0x14/Mythos-nano

Mythos-nano: A Compact Model for Advanced Reasoning

Mythos-nano, developed by void0x14, is a 3.1 billion parameter language model that distinguishes itself through its exceptional performance in competitive mathematics and coding challenges. Despite its small size, it achieves benchmark scores comparable to trillion-parameter systems, particularly when augmented with verifiable feedback mechanisms.

Key Capabilities & Performance

• Competitive Mathematics: Mythos-nano shows strong results on benchmarks like AIME25 (91.4%), AIME26 (94.3%), HMMT25 (89.3%), and BruMO25 (93.8%). With CLR (Cognitive Loop Reasoning), its performance significantly improves, reaching 96.7% on AIME25 and 97.1% on AIME26.
• Competitive Programming: The model excels at LeetCode-style problems, achieving a 96.1% aggregate pass-rate (123/128 problems), placing it very close to top-tier models like Gemini 3.1 Pro and GPT-5.3-Codex.
• Reasoning Focus: The core thesis behind Mythos-nano is that small models can achieve frontier reasoning capabilities through verifiable feedback.

Important Considerations

• Not for Tool-Calling: This model was not trained on tool-calling or agent-based programming data and is not recommended for tasks involving function calling or API orchestration.
• Abliterated (Uncensored): Mythos-nano has reduced safety guardrails, meaning it will not decline requests that a safety-tuned model normally would. Users are responsible for its outputs and legal compliance.

Usage Recommendations

• Sampling: Recommended temperature range is 0.6–1.0.
• Output Tokens: Supports up to 40960 output tokens for complex problems.
• Availability: GGUF formats (mythos-nano-f16.gguf, mythos-nano-Q4_K_M.gguf) are provided for llama.cpp and Ollama.