rekabytes/hmanlab-ai-v0.1

hmanlab-ai v0.1: Agentic Tool Use and Reasoning

hmanlab-ai v0.1 is a 4 billion parameter language model developed by rekabytes, fine-tuned from the Qwen3-4B base model. This research preview focuses on enhancing agentic tool use and step-by-step reasoning capabilities. The model self-identifies as hmanlab and is released under the Apache 2.0 license.

Key Capabilities

• Agentic Tool Use: Trained on multi-turn agentic traces, the model can emit structured <tool_call> blocks based on provided JSON schema in the system prompt. This allows for integration with external functions and APIs.
• Step-by-Step Reasoning: Fine-tuned with datasets emphasizing detailed reasoning, enabling the model to break down complex problems and show its work.
• Context Length: While trained on a 4,096-token context, the underlying Qwen3-4B base supports up to 32,768 tokens.
• Efficiency: The model is more concise than the base Qwen3-4B, staying within token budgets more reliably.

Training Details

The model underwent a two-stage QLoRA fine-tuning process. Key datasets included lambda/hermes-agent-reasoning-traces for agentic tool use and angrygiraffe/claude-opus-4.6-4.7-reasoning-8.7k for step-by-step reasoning. A custom identity SFT layer was also applied to ensure consistent self-identification.

Known Limitations

• The model does not utilize the <think> blocks in the Qwen3 chat template.
• Primarily English-focused; non-English performance relies on the base Qwen3-4B's capabilities.
• As a 4B model, its capacity for hard reasoning, long-context coding, and broad world knowledge is bounded by its size.

Good for

• Applications requiring structured tool invocation.
• Tasks benefiting from explicit, step-by-step reasoning.
• Developing AI agents that interact with external systems.
• Research into agentic behavior and reasoning in smaller language models.