zombiegirlcz/kali-assistant-1.5b

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The zombiegirlcz/kali-assistant-1.5b is a 1.5 billion parameter instruction-tuned agent model, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct, specifically designed for on-device operation within the Kali AI Assistant Android app. It excels at mapping natural language requests (Czech/English) to specific tool calls, including PRoot Linux commands, NetHunter CLI, Android host checks, and native phone actions. This model is optimized for precise tool selection and execution in a mobile environment, offering a 100% tool-name accuracy on held-out prompts.

Loading preview...

Kali Assistant 1.5B Overview

zombiegirlcz/kali-assistant-1.5b is a specialized 1.5 billion parameter agent model, fine-tuned from Qwen/Qwen2.5-1.5B-Instruct. Its primary purpose is to serve as an on-device assistant within the Android app Kali AI Assistant, translating natural language commands into actionable tool calls.

Key Capabilities

  • Tool-Oriented Agent: Designed to accurately map user requests to a predefined set of tools for mobile device interaction.
  • Extensive Toolset: Supports a wide range of actions including:
    • Executing PRoot Linux commands (proot_exec)
    • Performing read-only Android host checks (host_shell)
    • Accessing device state (battery, Wi-Fi, location, core status)
    • Controlling native phone actions (opening apps, URLs, dialing, sending SMS, playing media).
  • High Accuracy: Achieves 100% tool-name accuracy on 14 held-out prompts, significantly outperforming its base model (64%).
  • Multilingual Support: Trained on both Czech and English agent trajectories.
  • On-Device Optimization: Fine-tuned for efficient operation on mobile devices, with a merged fp16 model and GGUF quantization available.

Training Details

The model was fine-tuned using LoRA on a single T4 GPU, leveraging approximately 1.1k curated agent trajectories. These trajectories were deterministically generated from the app's command surface and verified device outputs, ensuring high relevance and accuracy for its intended use case.

Recommended Usage

For optimal performance and native tool calling, llama.cpp with the --jinja flag is recommended. The model can be seamlessly integrated into the Kali AI Assistant app by configuring the custom provider settings.

Finetunes
1 models