Is Needle 2 free for commercial use?

Yes. Needle 2 is licensed under Apache-2.0, with weights published on Hugging Face at Cactus-Compute/needle2 and source in the cactus-compute/needle repo. That covers commercial products, and Cactus points to Pebble as a shipping example — the Index 01 app runs Needle locally so a screenless ring can act on speech without a network.

How is Needle 2 different from the first Needle?

The first Needle, published in May 2026, was a 26M-parameter model distilled from Gemini tool calling. Needle 2 is a 45M-parameter model trained from scratch on Cactus's Simple Attention Network architecture, using a proprietary 115B-token pretraining corpus plus 38B post-training tokens, and it adds device use and structured extraction to plain function calling.

What hardware can actually run Needle 2?

Needle 2 reaches 500 tokens per second of decode on a Raspberry Pi 5, 400–1,500 on VR headsets such as Meta Quest 3S and Apple Vision Pro, and 300–700 on sub-$200 phones like Samsung's A-Series. The roughly 28MB peak session RAM means it also fits newer microcontrollers, including the ESP32-S3.

Can I fine-tune Needle 2 on my own product's tools?

Yes, and Cactus designed for it. At 45M parameters, Needle 2 is small enough to retrain on the machine you develop on: the GitHub repo and the cactus-needle Python package handle tuning and testing on a Mac or PC in minutes to a few hours, so you can ship a model that speaks your device's tool vocabulary.

What does Needle 2 do when a request is out of scope?

Needle 2 refuses by returning an empty call rather than guessing. Every response also carries a learned confidence score, so a product can set a threshold: act above it, and re-ask or escalate to a cloud model below it. Cactus argues most device requests are routine control, so escalation stays rare.

Cactus Compute · 2026-08-10 · major

Needle 2 — 14MB agentic model for phones, robots and microcontrollers

Needle 2 is a 45M-parameter open model for tool calling and structured extraction that ships as one 14MB binary and runs a full session in 28MB of RAM. It scores 63.7% on Mobile Actions under Apache-2.0.

Cactus Needle 2 announcement card for the 14MB agentic model

A 45M-parameter tool-calling model that ships as one 14MB binary and runs on hardware too small for anything else.

Key specs

Parameters	45M
File size	14 MB

Quick facts

Maker	Cactus Compute
License	Apache-2.0
Session RAM	28 MB peak
Quantization	CQ2-bit, trained from pretrain through post-train
Speed	500+ tok/s decode on a Raspberry Pi 5
Runs on	ARM64, x86-64, ARMv7, RISC-V and WebAssembly
Weights	huggingface.co/Cactus-Compute/needle2

Benchmarks

Mobile Actions (961 rows, ordered strict exact match)

Needle 2 (45M, CQ2-bit)		63.7%
LFM2.5 230M (f16, vLLM)		69.1%
Apple FM (~3B, on-device)		64.5%
FunctionGemma 270M (f16, vLLM)		64%

source ↗

What is it?

Needle 2 packs tool calling, device use, and structured extraction into a single 14MB binary that needs about 28MB of RAM for a whole session. Cactus Compute is aiming at the cheap end of the edge — budget phones, wearables, small robots, and microcontrollers — rather than the Macs and PCs that usually get called 'on-device AI'.

How does it work?

The architecture is Cactus's Simple Attention Network, an attention-only stack described in a July 2026 arXiv paper. A Hadamard MLP swaps dense projections for a fixed Walsh transform, hashed n-gram 'engram' tables hold world knowledge outside the layer stack, and a 256-token sliding window bounds the KV cache. Weights, activations, and cache all train at CQ2-bit precision, so the 2-bit model deployed is the model that was trained.

Why does it matter?

Tool calling does not need world knowledge — mapping a messy sentence onto a typed function is a much smaller problem than chat, and Needle 2 is the argument that 45M parameters cover it. On Mobile Actions it lands within a point and a half of a 3B Apple model and beats FunctionGemma 270M, at a fraction of the size. That puts an offline agent on hardware with no GPU, no NPU, and a few hundred MB of RAM.

Who is it for?

embedded and mobile developers, robotics teams

Frequently asked questions

Is Needle 2 free for commercial use?: Yes. Needle 2 is licensed under Apache-2.0, with weights published on Hugging Face at Cactus-Compute/needle2 and source in the cactus-compute/needle repo. That covers commercial products, and Cactus points to Pebble as a shipping example — the Index 01 app runs Needle locally so a screenless ring can act on speech without a network.
How is Needle 2 different from the first Needle?: The first Needle, published in May 2026, was a 26M-parameter model distilled from Gemini tool calling. Needle 2 is a 45M-parameter model trained from scratch on Cactus's Simple Attention Network architecture, using a proprietary 115B-token pretraining corpus plus 38B post-training tokens, and it adds device use and structured extraction to plain function calling.
What hardware can actually run Needle 2?: Needle 2 reaches 500 tokens per second of decode on a Raspberry Pi 5, 400–1,500 on VR headsets such as Meta Quest 3S and Apple Vision Pro, and 300–700 on sub-$200 phones like Samsung's A-Series. The roughly 28MB peak session RAM means it also fits newer microcontrollers, including the ESP32-S3.
Can I fine-tune Needle 2 on my own product's tools?: Yes, and Cactus designed for it. At 45M parameters, Needle 2 is small enough to retrain on the machine you develop on: the GitHub repo and the cactus-needle Python package handle tuning and testing on a Mac or PC in minutes to a few hours, so you can ship a model that speaks your device's tool vocabulary.
What does Needle 2 do when a request is out of scope?: Needle 2 refuses by returning an empty call rather than guessing. Every response also carries a learned confidence score, so a product can set a threshold: act above it, and re-ask or escalate to a cloud model below it. Cactus argues most device requests are routine control, so escalation stays rare.

Try it

pip install cactus-needle