squ11z1/Mythos-nano

Mythos-nano: A Small Model for Frontier Reasoning

Mythos-nano is an independent 3 billion parameter open model project developed by squ11z1, distinct from official Anthropic releases. Its core thesis is to achieve frontier reasoning capabilities in a small model size through verifiable feedback mechanisms.

Key Capabilities & Performance

• Exceptional Mathematical & Coding Performance: Mythos-nano demonstrates strong performance on competitive mathematics benchmarks (AIME, HMMT, BruMO) and competitive programming problems (LeetCode contests), often rivaling much larger, trillion-parameter models.
• Competitive Programming Focus: It is specifically recommended for tasks involving competitive programming problems (e.g., LeetCode-style challenges).
• Reasoning with CoT-L: When combined with CoT-L (Chain-of-Thought with Lookahead), Mythos-nano's performance significantly improves across mathematical benchmarks, showcasing its ability to leverage advanced reasoning techniques.

Important Considerations

• Not for Tool-Calling/Agents: This model was not trained on tool-calling or agent-based programming data and is therefore not recommended for function calling, API orchestration, or autonomous coding agents.
• Abliterated (Uncensored): The refusal direction has been removed, meaning safety guardrails are reduced. Users are solely responsible for outputs and legal compliance.

Usage

• Recommended sampling temperature: 0.6–1.0.
• Supports up to 40960 output tokens for complex problems.
• GGUF versions (mythos-nano-f16.gguf, mythos-nano-Q4_K_M.gguf) are provided for llama.cpp and Ollama.