karthik-2905/AL1-model-B

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

karthik-2905/AL1-model-B is a compact 0.6 billion parameter language model, fine-tuned from Qwen3-0.6B, specifically optimized for native tool-calling capabilities. It reliably emits Hermes-style tags for a small set of file and shell tools, achieving 91.7% accuracy on tool argument parsing. This model is designed for applications requiring efficient and accurate function calling within a small footprint, while maintaining strong general chat coherence.

Loading preview...

AL1-model-B: Compact Tool-Calling LLM

AL1-model-B is a 0.6 billion parameter language model developed by karthik-2905, fine-tuned from the Apache-2.0 licensed Qwen/Qwen3-0.6B base. Its primary innovation lies in its native tool-calling reliability, achieved through a LoRA supervised fine-tune (SFT) targeting Hermes-style <tool_call> emission.

Key Capabilities & Performance

  • Highly Accurate Tool-Calling: Achieves 100% tool name accuracy and 91.7% full argument accuracy (11/12 cases) on a held-out evaluation set, a significant improvement over the base model's 50% accuracy.
  • Compact Size: At 0.6B parameters, it offers efficient deployment for tool-use cases.
  • General Chat Coherence: Maintains strong general chat capabilities with no observed regression compared to its base model.
  • Supported Tools: Designed to interact with five specific file/shell tools: read_file, write_file, edit_file, list_dir, and run_bash.

Training & Development

The model was fine-tuned using a LoRA adapter on a dataset of 1,423 train and 74 validation examples, with a ~2:1 tool:chat ratio to enhance tool-calling while preserving chat quality. The adapter was then merged into the base model for a self-contained, standalone fp16 model.

Limitations

  • As a small model (0.6B), it is not a substitute for larger instruct models.
  • Chat quality and broad-knowledge regression (e.g., MMLU/GSM8K) are currently evaluated by eyeballing, not formal LLM-judge scoring or benchmarks.
  • 4-bit quantization has not been tested, and may lead to more significant degradation than typically seen in larger models.

Ideal Use Cases

  • Resource-constrained environments requiring reliable function calling.
  • Agentic workflows where precise tool invocation is critical.
  • Applications needing a small, specialized model for interacting with file systems or executing shell commands.

Finetunes
1 models