Cactus Compute Releases Needle 3: Ultra-Compact Model for On-Device Automation
Cactus Compute has released Needle 3, an 8–29MB sliceable 121M-parameter foundation model optimized specifically for on-device function calling and structured schema extraction. By completely removing general chat capabilities, the model achieves benchmark performance on device command tasks that matches or exceeds models up to 10 times its size, including DeepSeek V4 Flash. Needle 3 proves that lightweight, highly specialized architectures can perform local automation and tool execution entirely offline without requiring massive cloud APIs or high-RAM devices. This approach significantly reduces latency, compute costs, and memory footprints for mobile and edge AI applications while ensuring data privacy. The model replaces dense feed-forward layers with a Monarch Hadamard MLP and stores 70.8M parameters in hashed n-gram tables (engrams), allowing the 121M model to run at the arithmetic workload of a 50M parameter model (100 MFLOPs per token). Furthermore, it guarantees valid schema parsing by using byte-level grammars compiled directly from tool definitions to format JSON outputs reliably.
## BACKGROUND
Function calling enables large language models to translate natural language user requests into structured data payloads that software tools and system APIs can execute directly. Standard foundation models often require billions of parameters to handle general conversational context alongside function calling, making them heavy and slow for on-device deployment.