Michael-Kozu/Deimos-A1

Overview

Deimos A1 is a 4 billion parameter model developed by Michael-Kozu, representing the first public iteration (Alpha 1) of the Deimos line. It is a concise chain-of-thought (CCoT) fine-tune of the Qwen3.5-4B base model, specifically engineered to generate highly condensed and efficient reasoning traces.

Key Capabilities

• Concise Reasoning: Produces dense, stepwise <think> blocks that average ~1/8 the tokens of the base model, significantly reducing output length.
• Improved Efficiency: Achieves approximately 6 times faster wall-clock inference time on reasoning benchmarks compared to its base model due to reduced token generation.
• Enhanced Accuracy: Improves accuracy on measured reasoning benchmarks, despite the reduced token count in its thought process.
• Specialized Training: Fine-tuned on the 4,919-row Michael-Kozu/Quark CCoT SFT dataset, where reasoning traces were compressed by a Qwen3.6-35B teacher model.

Use Cases

• Efficient Reasoning Applications: Ideal for scenarios requiring quick and resource-light generation of logical thought processes.
• Text Generation with Structured Thinking: Suitable for tasks where a clear, condensed chain of thought is beneficial before generating a final answer.

Limitations

• Inherits limitations of the Qwen3.5-4B base model, including language coverage and knowledge cutoff.
• Currently English-only.
• Mild overfitting observed in the final training epoch, though a lower-validation checkpoint is preserved internally.