Is Needle 3 free to use in a commercial product?

Needle 3 is released under Apache-2.0, on both the GitHub repository and the Hugging Face weights repo. That permits commercial use, modification and redistribution as long as the licence and attribution notices are kept. Cactus publishes the weights, the needle3.safetensors fine-tuning checkpoint and the per-platform engines, so a product can ship the model without calling a Cactus service.

What hardware can Needle 3 actually run on?

Cactus targets Needle 3 at phones, wearables, smart home devices, robots, cars and microcontrollers, and the build tool produces engines for macOS, Linux, Windows, Android, iOS and the browser. On a Raspberry Pi 5 the published figures are 400-4k tokens per second of decode and 1-10k tokens per second of prefill, with the exact rate depending on which layer depth you deploy.

Can Needle 3 be fine-tuned on my own tasks?

Yes. Cactus ships a needle3.safetensors checkpoint alongside the deployable needle3.cact file specifically for fine-tuning. On the DroidCall benchmark, tuning lifted every subnetwork by 18-36 points, and Cactus reports that from 4 layers up the tuned subnetwork passes DeepSeek V4 Flash on that task, starting at 29M parameters.

Which benchmarks did Cactus test Needle 3 on?

Cactus reports Needle 3 results on three tool-calling and extraction sets: Mobile Actions (961 rows, scored on exact calls), DroidCall (200 rows, scored on exact calls in order) and BFCL v4 (3,641 rows, scored by AST match). The published claim from those runs is that Needle 3 beats models 10x its size on mobile tool calls and matches models 2-3x its size on extraction.

Cactus Compute · 2026-09-18 · major

Needle 3 — an 8-29 MB model that picks tools on a Raspberry Pi

Needle 3 is Cactus Compute's 8-29 MB automation model for phones, wearables and microcontrollers. Every depth from 2 to 20 layers of the same weights ships as its own deployable model. Apache-2.0.

Hugging Face model card for Cactus Compute Needle 3, an 8-29 MB on-device automation model

One Apache-2.0 weight file that can be cut to any depth from 2 to 20 layers and still run on a microcontroller.

Quick facts

Maker	Cactus Compute
License	Apache-2.0
Size on disk	8-29 MB (CQ2 2-bit)
Parameters	29M-121M, by layer depth
Deployable depths	Every depth from 2 to 20 layers
Speed (Raspberry Pi 5)	400-4k tok/s decode, 1-10k tok/s prefill
Availability	Weights and engines on Hugging Face, code on GitHub

What is it?

Needle 3 adds what Cactus calls intelligence laddering: every depth from 2 to 20 layers of the same network is a model you can deploy on its own. The full 20-layer version holds 121M parameters and ships as an 8-29 MB binary in Cactus's CQ2 2-bit format. Cactus Compute built it for three jobs on small hardware — tool calling, structured extraction, and text embeddings for local search and routing.

How does it work?

The architecture is a Laddered Simple Attention Network, so training produces one set of weights in which each depth from 2 to 20 layers works as a standalone model. A 2-layer cut is the smallest deployable subnetwork; the 20-layer version is the largest. Cactus says the 121M model does the arithmetic of a 50M one. Each target platform ships a prebuilt engine under 1 MB that loads the needle3.cact weight file at start, and a needle3.safetensors checkpoint is published for fine-tuning.

Why does it matter?

Keeping tool calls on the device removes the round trip to a cloud API, which matters for battery, latency and privacy on hardware that may be offline. Cactus reports the model beats models 10x its size on mobile tool calls and matches models 2-3x its size on extraction. Fine-tuning on DroidCall raises every subnetwork by 18-36 points, and from 4 layers up the tuned subnetwork passes DeepSeek V4 Flash on that task.

Who is it for?

mobile and embedded developers

Frequently asked questions

Is Needle 3 free to use in a commercial product?: Needle 3 is released under Apache-2.0, on both the GitHub repository and the Hugging Face weights repo. That permits commercial use, modification and redistribution as long as the licence and attribution notices are kept. Cactus publishes the weights, the needle3.safetensors fine-tuning checkpoint and the per-platform engines, so a product can ship the model without calling a Cactus service.
What hardware can Needle 3 actually run on?: Cactus targets Needle 3 at phones, wearables, smart home devices, robots, cars and microcontrollers, and the build tool produces engines for macOS, Linux, Windows, Android, iOS and the browser. On a Raspberry Pi 5 the published figures are 400-4k tokens per second of decode and 1-10k tokens per second of prefill, with the exact rate depending on which layer depth you deploy.
Can Needle 3 be fine-tuned on my own tasks?: Yes. Cactus ships a needle3.safetensors checkpoint alongside the deployable needle3.cact file specifically for fine-tuning. On the DroidCall benchmark, tuning lifted every subnetwork by 18-36 points, and Cactus reports that from 4 layers up the tuned subnetwork passes DeepSeek V4 Flash on that task, starting at 29M parameters.
Which benchmarks did Cactus test Needle 3 on?: Cactus reports Needle 3 results on three tool-calling and extraction sets: Mobile Actions (961 rows, scored on exact calls), DroidCall (200 rows, scored on exact calls in order) and BFCL v4 (3,641 rows, scored by AST match). The published claim from those runs is that Needle 3 beats models 10x its size on mobile tool calls and matches models 2-3x its size on extraction.

Try it

pip install cactus-needle