ClarkBear/gemma4-e2b-mobile-actions-200

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:5.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 24, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

ClarkBear/gemma4-e2b-mobile-actions-200 is a 5.1 billion parameter Gemma 4 E2B model fine-tuned for mobile action function calling. Developed by ClarkBear, this model converts natural language mobile assistant requests into specific tool calls, such as creating calendar events or showing maps. It was trained on 200 samples from the google/mobile-actions dataset and is optimized for on-device inference, with LiteRT-LM artifacts available for mobile deployment.

Loading preview...

Overview

This model, ClarkBear/gemma4-e2b-mobile-actions-200, is an experimental 5.1 billion parameter Gemma 4 E2B checkpoint. It has been fine-tuned specifically for mobile action function calling, converting natural language requests into structured tool calls. The training was performed locally on an Apple M1 Pro using LoRA, and the merged weights are provided for direct Transformers inference.

Key Capabilities

  • Natural Language to Tool Call Conversion: Translates user requests like "Show me Patisserie Valerie on Kensington High Street" into executable tool calls (e.g., call:show_map{query:"Patisserie Valerie at 208 Kensington High Street, London, W8 7RG"}).
  • Supported Functions: Handles a range of mobile actions including create_calendar_event, create_contact, show_map, open_wifi_settings, send_email, turn_on_flashlight, and turn_off_flashlight.
  • On-Device Deployment: Includes pre-converted LiteRT-LM artifacts (.litertlm files) for efficient deployment and experimentation on mobile devices (Android/iOS), with both 8-bit and 4-bit quantized versions available.

Performance and Training Details

Evaluated on 200 held-out examples from the google/mobile-actions dataset, the model achieved an 85.0% exact match rate, with a 94.0% format valid rate and 94.0% function name accuracy. It was trained on only 200 samples for 1 epoch, using google/gemma-4-E2B-it as the base model and a max sequence length of 1024 tokens.

Limitations

As an experimental fine-tune on a small dataset, the model has limitations:

  • May occasionally emit extra tool calls.
  • show_map accuracy is sensitive to address formatting.
  • Calendar event creation is more challenging due to the need for title and datetime extraction.
  • LiteRT-LM artifacts can be large (up to 5 GB), requiring sufficient device storage and memory.

Finetunes
2 models