Needle, released by Cactus Compute on May 15, 2026, is an ultra-lightweight, function-calling-only small language model with 26 million parameters, designed to operate at ultra-high speeds in extremely resource-constrained on-device edge environments such as smartphones, smartwatches, and AR glasses. Much like a smart device's hardware controller detects sensor interrupts and instantly converts signals with zero latency, Needle acts as a digital micro-router, mapping users' voice or text queries in real-time in a mobile environment and connecting them with available API tools in a single forward pass. It employs a unique, simplified attention network architecture that completely removes heavy multi-layer perceptron layers, achieving an extremely compact size while ensuring knowledge retrieval performance for single-query purposes. Existing general-purpose small models still pack billions of parameters, making them slow to run on edge devices, consuming significant device memory and power during computation, and limiting battery life. To overcome these limitations, Needle adopts a bold architecture optimization that eliminates multi-layer perceptron layers and incorporates a precise distillation technique from Gemini 3.1, enabling it to demonstrate on-device response speeds that are tens of times faster than much larger general-purpose models, using only a 14MB, 4-bit quantized model. It exhibits speeds capable of processing thousands of tokens of context in real-time, even on the low-power processors of smartphones or smart glasses, and achieves unparalleled practicality by performing complete tool calls solely on the device, without complex pipeline detours or the latency of cloud servers. This tool offers a highly innovative scenario for biotechnology researchers who need to use mobile tablets or wearable devices inside sterile laboratories where their hands are not free. For example, when a researcher performs pipetting or microscopy and instructs a wearable device in natural language, such as "Record the current humidifier temperature sensor value and check the comparison results of the gene expression sequence registered yesterday," Needle operates in the backend. Running on the embedded core within the device, Needle receives a list of local laboratory control functions and protein analysis API tool definitions as input and instantly generates a structured JSON message combined with the corresponding call parameter values with zero latency. This allows researchers to reliably operate a smart lab agent environment where they can intuitively control experimental equipment and analysis tools solely on the device, without the need for complex server operations or unstable network conditions.

Needle, released by Cactus Compute on May 15, 2026, is an ultra-lightweight, function-calling-only small language model with 26 million parameters, designed to operate at ultra-high speeds in extremely resource-constrained on-device edge environments such as smartphones, smartwatches, and AR glasses. Just as the hardware controller of a smart device detects a sensor interrupt and instantly converts the signal with zero latency, Needle is a digital micro-assistant that maps the user's voice or text queries in real-time in a mobile environment and connects them to available API tools with a single forward pass.

Natural Language Bio Tool Binding for Wearable Devices: In an offline lab environment, recognize a researcher's natural language commands on mobile eyewear and map the parameters for invoking a gene reference tool (e.g., protein_search) within 0.1 seconds for voice-activated execution.