Nemesispro/Mythos-nano

TEXT GENERATIONConcurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 5, 2026License:mitArchitecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Mythos-nano by Nemesispro is a 3.1 billion parameter causal language model with a 32768 token context length. This independent open model project excels in competitive programming and mathematics, demonstrating strong reasoning capabilities comparable to much larger models. It is specifically optimized for solving LeetCode-style problems and complex mathematical challenges, achieving high scores on benchmarks like AIME and HMMT. The model is uncensored, with safety guardrails reduced, and is not recommended for tool-calling or agent-based programming tasks.

Loading preview...

Mythos-nano: A Compact Model for Competitive Programming and Mathematics

Mythos-nano, developed by Nemesispro, is a 3.1 billion parameter language model designed for high-performance reasoning in competitive programming and mathematics. Despite its small size, it demonstrates capabilities often seen in trillion-parameter systems, a core thesis achieved through verifiable feedback mechanisms.

Key Capabilities & Performance

  • Exceptional Mathematical Reasoning: Achieves strong results on competitive mathematics benchmarks such as AIME (91.4%) and HMMT (89.3%), with a variant (Mythos-nano + CLR) reaching 96.7% on AIME25 and 97.1% on AIME26.
  • Competitive Programming Prowess: Excels at LeetCode-style problems, achieving a 96.1% pass-rate (123/128) in Python, placing it competitively against much larger models like Gemini 3.1 Pro and GPT-5.2.
  • Long Context Handling: Supports a context length of 32768 tokens, allowing for complex problem statements and extended interactions.

Important Considerations

  • Not for Tool-Calling: This model was not trained on tool-calling or agent-based programming data and is not recommended for function calling, API orchestration, or autonomous coding agents.
  • Abliterated (Uncensored): The model's refusal direction has been removed, meaning it will not decline requests a safety-tuned model normally would. Users are solely responsible for outputs and legal compliance.

Usage Recommendations

  • Optimal for: Competitive programming problems (e.g., LeetCode), mathematical problem-solving.
  • Recommended Sampling: Temperature between 0.6–1.0, with up to 40960 output tokens for challenging problems.

Finetunes
1 models