Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash

Needle 3 is a compact foundation model built for mobile, wearables, robots, smart home, automotive and microcontroller deployment: a single 8-29 MB CQ2-bit binary implementing a Laddered Simple Attention Network with subnetworks from 2 to 20 layers. It focuses on on-device tool calling, structured extraction and text embeddings rather than general chat. Given tool definitions it picks which tool to call and fills every argument (returning ordered calls or an empty list if no tool applies), and it decodes typed JSON reliably for invoices, bookings and other schemas. The Python API is minimal: decorate functions as tools, add regex triggers to force routing, or pass a Pydantic model to extract structured fields. Each turn returns a JSON object with function_calls, reasoning, a calibrated confidence score (with suppressed_calls below the floor), and runtime metrics (prefill/decode TPS, peak RAM).

Architecturally it ships a laddered model family (29-121M parameters, 360B tokens of training) that produces 9-29 MB binaries for different depths; inference runs at roughly 400-4k tokens/s decode and 1-10k tokens/s prefill on a Raspberry Pi 5. Fine-tuning on the Cactus platform (LoRA then build .cact) lets small subnetworks specialize: a tuned 4‑layer subnetwork matches DeepSeek V4 Flash after one epoch and fine-tuning lifts subnetworks 18-36 points on DroidCall. Reported benchmarks show Needle 3 beating models 10× its size on mobile tool calls and matching models 2-3× larger on extraction. The platform provides datasets, quantisation, evaluation and prebuilt runtimes under 1 MB per target for easy deployment.