온디바이스 AI 'Needle2': 14MB 초소형 모델 공개

45M Params 800+ tok/s Pi5 prefill 500+ tok/s Pi5 decode CQ2-bit Compression 14 MB File size 28 MB Session RAM Our Bet Bringing On-Device AI to <$200 Devices : Edge AI has lately meant Macs and PCs, but the edge is mostly cheap hardware: over 21 billion connected IoT devices against roughly 1.5 billion PCs, and in emerging markets most phones ship under $200. Count budget phones, Raspberry Pis, microcontrollers, wearables, small robots like Reachy Mini, and connected home devices, and roughly four in five edge devices cost under $200. That is the hardware Needle targets: no GPU, no NPU, a few hundred MB of RAM. Function Call & Device Use : Turning on a light does not need a frontier model. A watch, a home, a robot: each already exposes its abilities as functions with typed parameters, so the only hard part is mapping a messy sentence onto them: which function, with which values. Framed that way, the problem needs no world knowledge and no open-ended prose, which is why 45M parameters suffice where chat needs billions. That smaller formulation is the bet everything else follows from. Extraction & Structured Outputs : The schema is the interface, and the same formulation covers documents: a schema plus a paragraph returns typed fields, an enum field is a classifier, an array field collects a list in one call. We enforce this with a contract, not a convention: every turn is answered with a call envelope, the empty call is the refusal, and a byte-level grammar compiled from the declared schemas constrains every token. The grammar carries the syntax, so all 45M parameters go to choosing functions and grounding arguments in the user&#x27;s words. Edge-Cloud Collaboration : No small model covers everything, so Needle says so instead of guessing: every response carries a learned confidence score, and off-topic requests return the empty call. Above your threshold, act; below it, re-ask or escalate to the cloud. Most device requests are routine control, so escalation stays rare and the default path stays private, instant, and free. Lossless 2bit Quantization : Small models break under post-hoc quantization, so we never quantize post-hoc: Needle 2 trains against Cactus Quants from pretrain through post-train, weights, activations, and KV cache alike. The 2bit model you deploy is the model that was trained. That is what fits 45M parameters into 14MB with nothing lost on our battery. Co-designed Model & Inference : Every architectural choice was benchmarked on the target hardware before it earned its parameters, and the deliverable is the pair, not the weights: a single dependency-free C++ binary that probes the CPU at startup and picks its kernels, with the model, tokenizer, and grammar compiler sealed inside. One artifact runs from Cortex-M to x86 to WebAssembly. There is nothing to install and nothing to download. Fine-tune on your Mac/PC : Every product has its own tool vocabulary, and a 45M model is small enough to retrain where it runs: the repo and python package tune and test on your own computer in minutes to a few hours. Ship a Needle that speaks your device&#x27;s tools, not a generic assistant. Production Needle is production-ready for products that require a minimal RAM footprint, low latency, privacy, and offline reliability. Pebble - the pioneer of the modern wearable industry - runs it locally in the Index 01 app to turn spoken requests into actions without depending on a network connection. “ The Pebble Index Ring has no screen. So when you speak to it, the action just has to happen, every time, with or without internet connection. We run Cactus Needle locally in the app, instead of relying on the cloud. The model&#x27;s footprint is tiny and the performance never lets us down. Architecture Needle 2 is pretrained on a proprietary 115B-token corpus and post-trained on 38B tokens with compact reasoning traces and careful dataset distribution design. For scale: LFM2.5-230M was pretrained on 19 trillion tokens, roughly 120× Needle&#x27;s total, and the evaluation below shows the two trading wins. Each component exists to buy capability without buying bandwidth. The Hadamard MLP replaces the usual dense up-and-down projections with a fixed Walsh transform and learned diagonals, so the channel mixing that dominates a small model&#x27;s weight reads costs almost no parameters at all. The engram moves world knowledge out of the stack into hashed n-gram tables that are read a few rows per token: capacity that is nearly free at decode time, which matters on devices where every megabyte read from flash is latency and battery. The multi-lane residual streams give a 27-layer, 512-wide network the routing flexibility of a much wider one, at the cost of a few dot products per layer rather than more attention or MLP volume. The memory system is designed backwards from fixed-RAM devices. Attention uses a 256-token sliding window so the KV cache is bounded no matter how long a session runs, and the system prompt and tool declarations are pinned as permanent sinks so the one thing a tool-calling model must never forget—its tools—is structurally unable to be evicted. The cache itself is trained with QAT, and weights are stored in Cactus Quants at a mixed bits per weight averaging 2bit. The result is that quality decisions and deployment decisions stay decoupled: one trained model, specialized to whatever precision and window a target device can afford. The engine earns its speed from what it refuses to compute. Weights never decompress into RAM: the 2-bit codes are expanded inside vector registers, fused into integer dot products, so resident memory stays at blob size and the arithmetic path is int8 end to end—activations, KV cache, and the lane routing tables alike. The grammar is an optimization, not just a guarantee: because the matcher knows which tokens are legal before the logits exist, the engine computes output scores only for candidate rows, skipping up to 98% of the vocabulary projection on structural tokens, and skips it entirely on steps whose output is already forced. One universal binary probes the CPU at startup and self-selects its kernel tier—SDOT, NEON, AVX2, RISC-V vectors, wasm SIMD, or scalar—and the thread pool spins through the short serial sections of a token instead of sleeping, which alone nearly doubled decode. None of this changes a single output: every trick is either exact or validated token-for-token against the reference path. All of it is ultimately an energy argument. On device silicon, moving a byte out of flash or DRAM costs orders of magnitude more than a multiply-accumulate, so the budget that matters is FLOPs per token and bytes per token together. The architecture cuts the first: a conventional transformer of Needle&#x27;s width and depth spends 164 MFLOPs per token, and even one squeezed down to Needle&#x27;s parameter count spends 87, because every parameter it owns must be exercised through a matmul. Needle spends 70, and keeps a fifth of its parameters as gathered memory that costs no arithmetic at all. The binary cuts the second, as the engine section showed: nothing rematerializes, the arithmetic stays int8 end to end, and the grammar prunes compute outright, so decoding a token reads at most the 14MB blob once, and on structural tokens meaningfully less. This is what battery life is made of. Even on a high-end phone, an always-on assistant lives inside a power budget; every MFLOP is milliwatt-hours, and Needle spends 7× to 85× fewer of them per token than the models it is benchmarked against. Bounded session memory is what puts microcontrollers in reach. Because the sliding window caps state, Needle 2&#x27;s RAM is a deterministic 28MB ceiling, not a curve that grows with conversation length. That fits MCU-class parts with external RAM, such as ESP32-P4 with 32MB of PSRAM, or STM32H7 and NXP i.MX RT boards with SDRAM. The engine compiles single-threaded for bare metal and ships as a static library for Cortex-M4, M7, an