AI / ML
Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash
Henry from Cactus has released Needle 3, a 8-29MB automation model that matches DeepSeek V4 Flash performance on a narrow task with just 4 layers of finetuning. The model supports 8 languages, has a confidence score for each response, and can be run on multiple platforms. It features intelligence laddering, Monarch Hadamard MLP, and tool calls with structured JSON output.