Needle 3

Needle 3 is a compact foundation model from Cactus Compute aimed at phones, wearables, robots, smart-home devices, automotive systems and microcontrollers. It is built on a "laddered simple attention network": each layer forms a sub-network of increasing capacity, so a single set of weights can run with anywhere from 2 to 20 layers. The 2-bit (CQ2) binaries range from roughly 8–9 MB to 29 MB.

It is designed mainly for function (tool) calling and structured extraction. The developer describes tools through Python decorators, and the model chooses the call and fills in arguments. Optional regex triggers force routing to a given tool. Each response is a JSON object with a calibrated confidence score; calls below a 0.1 floor are withheld, and the application can use the score to act, confirm or refuse. Cactus says it beats models 10x its size on mobile tool-calling benchmarks and matches 2–3x larger models on extraction. These are vendor claims.

Access is through the cactus-compute Python package, with prebuilt engines under 1 MB for targets such as macOS and Linux (x86_64, arm64, armv7, riscv64). The model can be LoRA fine-tuned and exported as a 4-bit .cact file for any subnetwork. The paid Cactus Platform offers datasets, quantisation, evaluation and full-depth fine-tuning. Cactus says fine-tuned subnetworks of 4 layers and up pass DeepSeek V4 Flash on DroidCall. The model is specialised for narrow tasks and is not a general chat model.