Cactus Compute releases Needle 2 tool-calling model

Cactus Compute has unveiled Needle 2, an open-source 45-million-parameter model optimized for tool calling, device control, and structured data extraction. Packaged as a single 14MB binary, the model operates within a strict 28MB RAM footprint. It runs on Cactus Quants at CQ2-bit precision, sealed inside a custom C++ engine that requires no external runtime or downloads during inference. This architecture allows the model to achieve impressive decode speeds on low-power hardware, reaching 500 tokens per second on a Raspberry Pi 5, 400 to 1,500 tokens per second on the Meta Quest 3S and Apple Vision Pro, and 300 to 700 tokens per second on budget smartphones costing under $200.

The model is built on a Simple Attention Network featuring 27 layers and a width of 512. This architecture swaps the traditional feed-forward network for a Hadamard MLP, retains Grouped-Query Attention, and incorporates engram key-value memory. Pretrained on a 115-billion-token corpus with 38 billion post-training tokens, Needle 2 consumes just 70 MFLOPs per token with 35 million active parameters. In comparison, LFM2.5-230M (pretrained on 19 trillion tokens) consumes 460 MFLOPs, FunctionGemma 270M uses 540 MFLOPs, and Apple FM requires roughly 6,000 MFLOPs.

For developers shipping firmware or apps on constrained hardware, Needle 2 offers a lightweight alternative for offline, on-device tasks. It ships as prebuilt binaries and a static library for macOS, Linux, Windows, Android, iOS, watchOS, tvOS, and WebAssembly, and is already used by Pebble in its Index 01 app. The model uses a 256-token sliding window to keep memory consumption flat, and constrains outputs using a byte-level grammar compiled from JSON schemas. In benchmark tests, Needle 2 scored 63.7 on Mobile Actions (961), trailing LFM2.5 230M at 69.1 and FunctionGemma 270M at 64.0, but beating Apple FM at 57.6. On DroidCall (200), it scored 17.0, matching FunctionGemma's 17.5 and beating LFM2.5's 11.0. It led both Seal-Tools splits, scoring 32.6 in-domain and 28.7 out-of-domain. On the BFCL v4 single-turn benchmark, it scored 42.6 overall, compared to 60.8 for LFM2.5 and 61.7 for Apple FM, while maintaining a 93.4% well-formed output rate.