Skip to content

Needle 2 Review

Needle 2 is Cactus Compute's free 45M-parameter model for on-device tool calling and structured extraction: a 14 MB binary that runs in about 28 MB of RAM. Needle 3 replaced it in September 2026.

Freemium · API

Stork's verdict on Needle 2

Needle 2 shows how far a 14 MB tool-calling model can go on cheap hardware, but new projects should start with Needle 3 and test accuracy on their own tools.

Needle 2 reviewed by Stork AI · stork.ai/en/needle-2

stork
I am a painting.
One still painting and one typed sentence.
This painting was never filmed.
A photo, a cartoon, a painting or a 3D character. From 10 cents a minute. Free to try.
Make a photo talkDevelopers: there is an API

What Needle 2 does

Needle 2 is a small language model from Cactus Compute built for one job: turning a plain-language request into the right function call, or pulling typed fields out of text. Cactus released it on August 9, 2026 as a 45M-parameter model compressed to about 2 bits per weight and baked into its own engine, so the whole thing is a single 14 MB binary with peak session RAM around 28 MB. It does not chat. You declare tools (a Python function with a docstring, or a JSON schema), and Needle picks which to call and fills the arguments; a byte-level grammar compiled from your schemas constrains every token, so output always parses. Each answer carries a calibrated confidence score, so an app can act above a threshold and escalate (for example to a cloud model) below it. A retrieval head lets you declare a large tool catalogue while only the top five tools are shown to the model per turn.

Cactus says Needle 2 decodes about 500 tokens a second on a Raspberry Pi 5, 300 to 700 on sub-$200 phones such as the Samsung A-Series, and runs on ARM64, x86-64, ARMv7, RISC-V and WebAssembly across Apple, Windows, Linux, Android and Raspberry Pi, down to microcontrollers such as the ESP32. You install it with pip install cactus-needle. On September 17, 2026 Cactus released Needle 3, an 8 to 29 MB model (29M to 121M parameters, any depth from 2 to 20 layers) that adds text embeddings and pairs with Cactus's Whistle speech model; the current package runs Needle 3 by default and keeps Needle 2 available with generation=2 for existing deployments.

Needle 2 pricing

OptionPriceIncluded
Needle 2 weights and engineFree, Apache License 2.0Model binary on Hugging Face, Python package cactus-needle, source and training code on GitHub
Needle 3 (current model)Free, Apache License 2.0Same package; weights, a fine-tunable checkpoint and per-platform engines on Hugging Face
Custom fine-tuning or hybrid checkpoints from CactusPricing not published; contact salesCactus offers fine-tuning and on-demand post-training through a "Talk to us" contact

There is no per-call fee: the model runs on your own device, offline. Cactus does not publish a pricing page.

Prices checked on the vendor's site on November 3, 2026. See cactuscompute.com and the Needle 2 model card.

Key features of Needle 2

What Cactus's Needle 2 release post and model card list:

  • 45M parameters at about 2 bits per weight (CQ2-bit), 14 MB binary, about 28 MB session RAM
  • Self-contained: model baked into the binary, no runtime downloads and no network needed
  • Tool calling and structured extraction, returned as JSON constrained by a grammar compiled from your schemas
  • Calibrated confidence score on every response for act-or-escalate routing
  • Tool retrieval: large catalogues, top five tools rendered per turn
  • 256-token sliding window with tools pinned, so memory stays flat in long sessions
  • Runs on ARM64, x86-64, ARMv7, RISC-V and WebAssembly; Python package cactus-needle
  • Browser sandbox to try example tools before integrating

Who Needle 2 is best for

Needle suits developers building voice or text control into cheap or offline hardware: smart home devices, wearables, robots, budget phones, Raspberry Pi projects and microcontrollers, where a general LLM will not fit and the task is mapping a sentence onto a known set of functions. It is not a chatbot and has no general knowledge to speak of. For new projects, start with Needle 3, which Cactus now recommends and which runs from the same package; use Needle 2 only if you already ship it and want to keep behaviour unchanged.

If you need a somewhat larger general-purpose small model, look at Google's Gemma family (including FunctionGemma, which Cactus benchmarks against). For running your own PyTorch models on phones, see ExecuTorch; for commercial on-device voice with wake words and intent recognition, Picovoice.

What users say about Needle 2

Reaction was mixed. A Developers Digest write-up of the Show HN launch says commenters liked the size and the in-browser demo, but posted confident wrong calls from the demo (for example a request to make a room warmer setting the thermostat to cool) and questioned how the benchmarks were framed. On Cactus's own figures cited there, Needle 2 scored 42.6 on BFCL v4 single-turn against 46.1 for FunctionGemma 270M, 60.8 for LFM2.5 230M and 61.7 for Apple FM, while running at 2 bits against their 16-bit weights. A Raspberry Pi blog post, written by the Cactus Compute team, shows it controlling LEDs and reading the CPU temperature on a Pi 5 in roughly 80 to 110 ms per call. Test it on your own tool set before relying on it.

Pros

  • Tiny: 14 MB binary, about 28 MB RAM, runs on CPU only
  • Free Apache-2.0 weights; runs fully offline
  • Output always parses thanks to schema-compiled grammar
  • Confidence score makes cloud fallback easy to wire
  • Runs on phones, Pis, browsers (WebAssembly) and microcontrollers

Cons

  • Superseded by Needle 3; kept mainly for existing deployments
  • Scores below larger small models on BFCL v4 single-turn
  • Launch testers reported confident wrong tool calls
  • No general chat or world knowledge

Needle 2 FAQ

What is Needle 2?

A 45M-parameter model from Cactus Compute for on-device tool calling, device control and structured extraction. It ships as a single 14 MB binary that runs in about 28 MB of RAM, with no network needed, and returns grammar-constrained JSON plus a confidence score.

Is Needle 2 free?

Yes. The Needle 2 model card on Hugging Face lists the Apache License 2.0, and the code is on GitHub. There is no per-call fee because it runs on your own device. Cactus's custom fine-tuning work has no published price.

What is the difference between Needle 2 and Needle 3?

Needle 3, released September 17, 2026, is an 8 to 29 MB model (29M to 121M parameters, selectable depth from 2 to 20 layers) that adds text embeddings and works with the Whistle speech model. The same cactus-needle package runs Needle 3 by default; pass generation=2 to keep Needle 2.

What devices can run Needle 2?

Cactus lists ARM64, x86-64, ARMv7, RISC-V and WebAssembly builds on Apple, Windows, Linux, Android and Raspberry Pi. It cites about 500 tokens a second decode on a Raspberry Pi 5, 300 to 700 on sub-$200 phones, and microcontrollers such as the ESP32.

Can Needle 2 chat or answer questions?

No. Cactus deliberately trades general chat ability for tool calling and extraction. It maps a request onto functions you declare, and returns an empty list for off-topic requests rather than guessing.

How accurate is Needle 2?

Cactus says it trades wins with FunctionGemma 270M, LFM2.5 230M and Apple FM on tool-calling benchmarks while being 5x to 70x smaller. On BFCL v4 single-turn it scored 42.6, below those three. Launch testers on Hacker News reported confident wrong calls, so validate on your own tools.

Compare Alternatives

See every Needle 2 alternative, compared

Needle 2

Freemium · API

From the Stork blog

Enjoying this? Get one like it in your inbox each morning.

one email a day · unsubscribe in two clicks · no third-party tracking