Michael-Kozu/Deimos-A4

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:May 3, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Deimos-A4 by Michael-Kozu is a 4.66 billion parameter causal language model built on Qwen3.5-4B, specialized for complex reasoning and hard mathematical tasks. It features an internal terse, concise 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 significant accuracy gains on benchmarks like leaderboard_math_hard and minerva_math500.

Loading preview...

Overview

Michael-Kozu's Deimos-A4 is a 4.66 billion parameter model, fine-tuned from Qwen3.5-4B, specifically designed for advanced reasoning and complex mathematical problems. It employs a unique internal "thinking" mechanism, generating compact, fragment-style reasoning within <think>...</think> blocks before producing a clean, professional user-facing response. This approach leads to substantial efficiency gains, with approximately 60% fewer tokens and 36% faster inference compared to the base Qwen3.5-4B model, while significantly improving accuracy on challenging math benchmarks.

Key Capabilities & Differentiators

  • Concise Reasoning: Internally generates terse, compact chain-of-thought, which is then expanded into a polished output, never exposing the internal fragments to the user.
  • Enhanced Math Performance: Achieves an average of +40 points accuracy on hard math tasks (AIME, MATH-hard, MATH-500) compared to its base model.
  • Token and Speed Efficiency: Demonstrates up to 77% token reduction and 36-47% faster inference times on hard math problems.
  • Specialized Training: Trained via length-biased rejection sampling on 4,338 "shortest-correct" traces, teaching the model to optimize for logical structure while minimizing verbosity.

When to Use Deimos-A4

  • Hard Math & Reasoning: Ideal for multi-step proofs, long algebraic chains, and complex problem-solving where the base model's reasoning capacity might be insufficient.
  • Code Generation: Effective for tasks requiring structured logical thought.
  • Optimized Inference: When token efficiency and faster response times are critical for reasoning-heavy applications.

Limitations

  • Not a Generalist: Shows knowledge regression on general knowledge recall (MMLU) and strict instruction-following tasks (IFEval) compared to the base model. For these, the base Qwen3.5-4B is recommended.
  • Internal Thinking Only: The model's compact reasoning style is designed to remain internal; stripping the chat template or forcing generation outside the <think> block may lead to suboptimal results.
  • 4B Pretraining Ceiling: Its underlying knowledge is limited by the Qwen3.5-4B base, focusing on reasoning depth rather than broad knowledge expansion.