AI Briefingneedle3: Tool calling, extraction, and embedding processing on mobile with an 8MB file
Cactus-Compute/needle3
About the project
A lightweight foundation model designed to run on constrained hardware environments such as mobile, wearable, and IoT devices. The entire model is packaged as a single file ranging from 8 to 29MB, performing tool calling, structured data extraction, and text embedding generation on-device without cloud connectivity.
It analyzes user requests to select the appropriate function exposed by the app and populate its arguments. Outputs are strictly parsed according to a defined grammar, ensuring no format errors, and each response includes a confidence score. The same model can also generate sentence vectors for local search and routing.
Applying the Laddered Simple Attention Network architecture, it supports subnetworks of varying depths from 2 to 20 layers. Through LoRA-based fine-tuning, small models optimized for specific product tools can be created, achieving performance comparable to DeepSeek V4 Flash starting from 29M parameters. Released under the Apache 2.0 license, it can be freely integrated into commercial projects.
Cactus-Compute/needle3
The original page has no description.
text-generation
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.