Cactus Compute has released Needle 2, an open model designed for tool calling, device control and structured data extraction on low-cost hardware. The 45-million-parameter model ships with its engine, tokenizer and grammar compiler in a single 14MB C++ binary and uses about 28MB of session memory.
The project deliberately avoids general chat. It maps natural-language instructions to functions and typed arguments exposed by a watch, home appliance, robot or application. Structured extraction uses the same mechanism: a supplied schema becomes a grammar that limits output to valid fields, enumerations, arrays and objects, preventing malformed structures.
Every response includes a learned confidence score. Product developers can set thresholds to execute locally, ask the user for clarification or escalate to a cloud model. Off-topic requests can return an empty call instead of a guess. That design positions Needle as one component in a hybrid system rather than a replacement for a broad language model.
Needle 2 uses a 256-token sliding attention window to keep memory bounded, while system instructions and tool declarations remain pinned so they are not evicted during a longer session. Weights, activations and the key-value cache were trained with Cactus's quantization approach instead of being compressed only after training. The deployed representation averages about two bits per weight.
The universal binary probes the processor and chooses among SDOT, NEON, AVX2, RISC-V vector, WebAssembly SIMD or scalar kernels. Cactus reports decode speed of about 500 tokens per second on Raspberry Pi 5, 400 to 1,500 on selected VR hardware and 300 to 700 on sub-$200 Samsung A-series phones. Newer microcontrollers with external memory are also listed as targets.
On Google's 961-row Mobile Actions evaluation, Needle achieved 63.7 percent ordered exact-match accuracy, behind LFM2.5 230M at 69.1 percent and FunctionGemma 270M at 64.0 percent, but ahead of an on-device Apple model at 57.6 percent in the published comparison. On two Seal-Tools tests, Needle led the listed small-model baselines. Cactus notes an important asymmetry: the competitors ran at 16-bit precision and were general models, while Needle is heavily compressed and narrowly trained for device actions.
The model is pretrained on a proprietary 115-billion-token corpus and post-trained on 38 billion tokens. It can be fine-tuned locally for a product's fixed tool vocabulary, and is licensed under Apache 2.0. Pebble uses it in the Index 01 app for offline spoken actions.
Needle 2's claim is therefore specialization under strict memory and power limits. Its own benchmarks show both competitive areas and clear deficits, making it most relevant where private offline function selection matters more than open-ended knowledge or conversation.
