0xAbhi/qwen3-0.6b-rc-car

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 5, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

0xAbhi/qwen3-0.6b-rc-car is a 0.8 billion parameter Qwen3-based model fine-tuned by 0xAbhi for translating natural language RC car driving commands into structured JSON tool calls. With a 32768 token context length, it specializes in generating specific control sequences for microcontrollers, such as 'Forward', 'Turn_Left', and 'Stop'. This model is optimized for direct JSON output, making it ideal for embedded systems and robotics control applications.

Loading preview...

Model Overview

0xAbhi/qwen3-0.6b-rc-car is a specialized 0.8 billion parameter model built upon the Qwen3-0.6B architecture. Its core function is to convert plain English instructions for controlling an RC car into a structured JSON array of tool calls. This model is specifically designed for direct machine interpretation, making it suitable for integration with microcontrollers or other robotic control systems.

Key Capabilities

  • Natural Language to Tool Call Translation: Converts commands like "go forward for 5 seconds then turn left" into a JSON array of actions, e.g., [{"name":"Forward","args":{"duration":5}},{"name":"Turn_Left","args":{}},{"name":"Stop","args":{}}].
  • Defined Tool Schema: Recognizes five specific tools: Forward, Backward, Turn_Left, Turn_Right, and Stop, each with predefined arguments.
  • Intelligent Command Interpretation: Automatically handles unspecified durations (defaults to 2 seconds), converts word-numbers to digits, clamps durations to a 1-10 second range, and ensures every output sequence ends with a single Stop command.
  • Optimized for Direct JSON Output: Fine-tuned to produce JSON directly without intermediate reasoning blocks, requiring enable_thinking=False when using the chat template.

Training Details

The model was fine-tuned using QLoRA (4-bit) on unsloth/Qwen3-0.6B-unsloth-bnb-4bit as its base. Training utilized a custom dataset, 0xAbhi/rc-car-commands, comprising 700 hand-authored command-to-tool-call JSON pairs. This specialized dataset ensures high accuracy within its narrow domain.

Limitations

  • Narrow Domain: Exclusively recognizes the five defined RC car control tools; it is not a general-purpose language model.
  • Language and Phrasing Specificity: Performance is optimized for English commands and phrasing patterns similar to its training data. Different phrasing styles or other languages may not yield correct translations.
  • Stylized Shapes: Composite commands for shapes (e.g., square, circle) result in stylized approximations based on fixed 90° turns, reflecting the training data's authoring rather than geometric precision.
  • Integer Durations: All durations are integers clamped between 1 and 10 seconds.

Finetunes
1 models