Cactus debuts Needle 2, a 14MB agentic LLM for edge devices

Cactus has launched Needle 2, an Apache 2.0-licensed, 45-million-parameter model that compresses down to a 14MB binary. Designed for edge computing, the model targets hardware under $200, such as budget smartphones, Raspberry Pis, microcontrollers, and wearables like the Pebble Index Ring. It operates without a GPU or NPU, requiring only a few hundred megabytes of RAM and establishing a deterministic 28MB RAM ceiling. This allows it to run on microcontrollers like the ESP32-P4 with 32MB of PSRAM, STM32H7, and NXP i.MX RT boards.

The model achieves its small footprint through a Simple Attention Network architecture and 2-bit quantization via Cactus Quants, which is integrated directly into training. Needle 2 was pretrained on a 115-billion-token corpus and post-trained on 38 billion tokens. It features a 256-token sliding window to bound the key-value cache. Needle 2 requires only 70 million active floating-point operations (MFLOPs) per token, compared to 164 MFLOPs for a dense transformer of the same shape, 460 MFLOPs for LFM2.5-230M, 540 MFLOPs for FunctionGemma-270M, and roughly 6,000 MFLOPs for Apple FM.

In evaluations, the 2-bit Needle 2 model proved competitive with larger, unquantized baselines. On the Google Mobile Actions benchmark, Needle 2 achieved 63.7% accuracy, trailing LFM2.5 at 69.1% and FunctionGemma at 64.0%, while outperforming Apple FM's 57.6%. On DroidCall, Needle 2 scored 17.0%, beating LFM2.5's 11.0%. It led on the Seal-Tools in-domain test with 32.6% accuracy and the out-of-domain test with 28.7% accuracy. On the Berkeley Function Calling Benchmark (BFCL) v4, it achieved an overall score of 42.6% and a 93.4% well-formed rate.

For developers, Needle 2 enables local, private, and zero-latency function calling on constrained hardware. Instead of relying on cloud APIs, practitioners can deploy a single, dependency-free C++ binary that runs across platforms from Cortex-M to WebAssembly. The model can be fine-tuned on a standard Mac or PC in a few hours, allowing developers to easily customize the model to their specific device schemas and tool vocabularies.