AI BriefingCactus Needle (GitHub Repository)
Key point
Needle, an ultra-lightweight AI model with just 26M parameters distilled from Gemini 3.1, has been released.
Details
Cactus Compute distilled Gemini 3.1 to develop Needle, a Simple Attention Network with just 26M parameters. This model enables local fine-tuning on Mac or PC, and achieves a prefill speed of 6,000 tokens per second and a decode speed of 1,200 tokens per second when run in the Cactus environment.
Needle was designed with Tiny AI for consumer devices such as smartphones, watches, and glasses in mind. In terms of single-shot Function Call performance, it outperforms models like FunctionGemma-270m and Qwen-0.6B, but it is optimized for tool calling in personal AI agents rather than conversational tasks.
Key technical features are as follows.
- Pre-training was conducted with 200B tokens using 16 TPU v6e.
- Post-training was completed with a single-shot function calling dataset of 2B tokens.
- The weights and data generation methodology are fully open via Hugging Face.
Users can test the model against their own tools through the provided Playground UI, and fine-tune it easily with a single click.
This summary 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 summary errors, attribution issues, or removal requests via Contact.
Related
Fake Customer Service Pages Induce ChatGPT and Gemini to Display Scam Numbers
Google Launches First Suncatcher Orbital AI Data Center Test on October 1
The Normalization of AI Translation: Principles of LLM-Based Translation and Structural Differences from NMT