Needle 26M open source model: Distilling Gemini tool calling capabilities, mobile phones can also run AI Agent

Core conclusion

The open source model Needle released by Cactus Compute achieves tool calling capabilities comparable to large models with only 26M parameters, and can run on terminal devices such as mobile phones, watches, and glasses. It is distilled from the Tool Calling capability of Gemini 3.1 and beats mainstream models such as FunctionGemma-270M and Qwen-0.6B on the single function call benchmark. For practitioners who make money from AI, this means that zero-cost deployment of AI Agents and automated workflows on edge devices has become a reality.

Background

In May 2026, Cactus Compute released the Needle model on GitHub under the MIT Open Source License and uploaded it to Hugging Face (model name: Cactus-Compute/needle). After the project was released, it quickly received 1948+ GitHub Stars and 96 Forks.

In terms of training, Needle used 16 TPU v6e to pre-train 200B tokens (which took 27 hours), and then used **a single function call data set of 2B tokens for post-training (only 45 minutes). The inference efficiency is amazing: Prefill speed is 6000 tokens/s, Decode speed is 1200 tokens/s, far exceeding most models of the same size on the market.

Needle Core Highlights

Implications for AI moneymakers

**The emergence of Needle directly reduces the deployment cost of AI Agent. ** In the past, running an AI model that could call tools required at least a few hundred MB of video memory. Now a 26M parameter model can accomplish the same task, and it is completely free.

Three clear directions for making money:

1. Localized AI Toolchain — Embed Needle into your own automation scripts to implement Function Calling without API Key, saving OpenAI API costs. To build a customized Agent process with n8n or Dify, please see our AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes.

2. Device-side AI product — Package Needle into a mobile application or wearable device for offline tool calling. For example, voice assistants on watches and real-time information retrieval on smart glasses. No Internet connection or cloud server is required, which is a huge bonus for developing independent products. See our previous coverage of Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo.

3. Low-cost fine-tuning vertical model - You can fine-tune Needle on your own MacBook and customize the exclusive API call Agent. Combined with How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial, an M4 Mac can run a complete AI automation workflow.

**Summary in one sentence: When AI can be put in your pocket and there is no need to pay to call Gemini or OpenAI, the door to the commercialization of tool-calling Agents will be completely open. **

Internal link guidance

• Want to know how to fully run an AI model on your local device? See our How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial for step-by-step instructions on how to configure a Mac/PC environment using LM Studio.
• For an in-depth understanding of the technical details and practical skills of AI Agent tool calling, it is recommended to read AI Agent Tools 2026 Complete Tutorial: 5 Tools to Build an Automated Pipeline in 30 Minutes.
• If you’re exploring using AI tools to make money on the side, check out how this developer implemented Indie Developer: n8n + OpenClaw Automation Workflow Earning $5,000/mo with the help of an automation platform.
• How can AI Agent help you make your first pot of gold? Reference OpenClaw AI Agent Playbook: 500K TikTok Views in 5 Days, $588 MRR.