Michael-Kozu/Deimos-A4
Deimos-A4 by Michael-Kozu is a 4.66 billion parameter causal language model built on Qwen3.5-4B, specialized for concise reasoning and hard mathematical tasks. It features an internal terse chain-of-thought mechanism, resulting in approximately 60% fewer tokens and 36% faster inference compared to its base model. This model excels in multi-step proofs and long algebraic chains, achieving over 40 points higher average accuracy on hard math benchmarks.
Loading preview...
Overview
Michael-Kozu's Deimos-A4 is a 4.66 billion parameter language model, fine-tuned from Qwen3.5-4B, specifically designed for complex reasoning and mathematical problem-solving. Its core innovation lies in an internal "terse, concise" chain-of-thought process, where the model generates compact reasoning within <think>...</think> blocks before producing a polished, user-facing response. This approach leads to significant efficiency gains, with approximately 60% fewer tokens and 36% faster inference compared to the base Qwen3.5-4B model, while also boosting accuracy on challenging math problems by over 40 points.
Key Capabilities
- Optimized for Hard Math: Achieves substantial accuracy improvements on benchmarks like AIME, MATH-hard, and minerva_math500, making it suitable for multi-step proofs and algebraic chains.
- Efficient Reasoning: Employs a unique internal concise reasoning mechanism, reducing token usage and accelerating inference speed.
- Configurable Thinking Mode: Features an
enable_thinkingtoggle in its chat template, allowing users to control the generation of internal reasoning traces based on task requirements.
Use Cases
- Hard Math & Reasoning: Ideal for problems requiring deep, multi-step logical deduction, such as advanced algebra, geometry, and number theory.
- Code Generation: While not explicitly detailed, the training pipeline includes code prompts, suggesting potential for code-related reasoning tasks.
Limitations
- Specialist Model: Deimos-A4 is a specialist and not a general-purpose replacement for its base model. It shows some knowledge regression on general knowledge (MMLU) and strict instruction-following tasks (IFEval) where its concise reasoning style can conflict with rigid formatting rules.
- Internal Reasoning: The compact reasoning fragments are intended to remain internal; stripping the chat template or forcing generation outside the
<think>block may lead to suboptimal results.
Training Details
Deimos-A4 was trained via length-biased rejection sampling on 4,338 verified "shortest-correct" traces, curated from its predecessor, Deimos-A1. This process taught the model to eliminate filler while maintaining logical structure. The model is released under the Apache 2.0 License.