[AI Starts to Abandon Chatting? Needle 3 Focuses Solely on Tool Invocation, Runs on as Little as 9MB] Beating AI News Flash: YC S25 edge AI company Cactus Compute has launched Needle 3, which specializes in tool invocation and structured extraction, without handling general chatting. For example, if a user says, 'Set the living room lights to 30%,' it only needs to locate the corresponding function and fill 'living room' and '30%' into the parameters. It is somewhat similar to the recently popular Jev, but follows a different approach. Jev directly scores candidate decisions, while Needle 3 still generates token by token, but its output is restricted to tool invocations that can be directly executed by programs. Needle 3's weight set can be trimmed into 2 to 20 layers. The 2-layer version is approximately 9MB, while the full 20-layer version is about 35MB, allowing users to choose based on the computational power of devices like smartwatches, Raspberry Pi, or smartphones. After specific fine-tuning for Android operational commands, the 4-layer, 29-million-parameter version achieved a self-test score of 62.5, surpassing DeepSeek V4 Flash's 60.5. [Original Link]