AI is collectively abandoning chat? 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 released Needle 3, specifically designed for tool calling and structured extraction, not for general chat. For example, if a user says "set the living room light to 30%," it only needs to find the corresponding function and fill in "living room" and "30%" as parameters.

It is somewhat similar to the recently popular Jev, but takes a different approach. Jev directly scores candidate decisions; Needle 3 still generates token by token, but its output is constrained to tool calls that programs can directly execute.

A single set of Needle 3 weights can be trimmed to 2 to 20 layers. The 2-layer version is about 9MB, and the full 20-layer version is about 35MB, allowing selection based on the compute power of different devices such as watches, Raspberry Pis, and phones.

After separate fine-tuning for Android operation commands, the 4-layer, 29-million-parameter version scored 62.5 in self-testing, surpassing DeepSeek V4 Flash's 60.5.