Skip to content
TrustList
N3
AI Model

Needle 3

New· 4

An 8 to 29 MB model that runs tool calls and structured extraction on phones, wearables and microcontrollers

About Needle 3

Most small language models try to be a little of everything. Needle 3, released by Cactus Compute on 18 September 2026, goes the other way: it is an automation foundation model built to run entirely on the device, and it gives up general conversation on purpose.

One network, many sizes

The full network has 121 million parameters. Cactus packs it into one file between 8 and 29 MB using its own 2-bit format, called CQ2. Any depth from 2 to 20 layers can also be deployed alone, so one release covers hardware from a microcontroller to a laptop. Cactus names a wide set of targets: phones, wearables, AR glasses, smart-home gear, robots, cars, Macs, PCs, game consoles, televisions and microcontrollers. It quotes decoding speeds as high as 4,000 tokens per second.

Three jobs, no chat

Needle 3 does tool calls, structured extraction and text embeddings. The company's founder has said publicly that packing conversational ability into a model this small is the part the team chose to give up. The weights are published on Hugging Face under Apache-2.0. Since inference happens on the device, nothing is billed per token, and Cactus treats that as the point of the design.

Benchmarks and the dispute around them

Cactus reports results on six benchmarks: Mobile Actions, DroidCall, BFCL v4, DSTC8 and two SNIPS sets, covering exact tool calls and field-level extraction on hundreds to thousands of rows each. It also says fine-tuning lifted each subnetwork by between 18 and 36 points on DroidCall. Those numbers appear only inside a chart image, so this description records what each benchmark measures and not scores.

The launch claim that Needle 3 matches DeepSeek V4 Flash is contested. A commenter on the vendor's launch thread reported far fewer correct tool shapes than another small model on their own test, several testers said the demo picked the wrong tool on plain requests, and others argued that a tool-calling model should not be set against DeepSeek V4 Flash at all. Two secondary write-ups date the release a day earlier; the 18th is used because the vendor's blog and launch thread agree on it.

TrustList summary

Needle 3 is an automation model from Cactus Compute, released on 18 September 2026 to run entirely on the device rather than behind an API. It is small enough to ship as a single file of 8 to 29 MB, and the same network can be deployed at several depths, so it fits a microcontroller as well as a laptop.

Main features

  • Tool calls: picks the tool an agent should call and fills in its arguments.
  • Structured extraction: pulls structured fields out of text.
  • Text embeddings: produces embeddings of text for matching and search.
  • Local running: runs without a call to an external service, on phones, wearables, cars, robots and microcontrollers.

Who it serves

Developers building on-device features for phones, wearables, smart homes, robots and other hardware the company lists.

Key facts

  • Size: 121M parameters, in the company's CQ2 2-bit format.
  • Licence: Apache-2.0, with weights published on Hugging Face.
  • Decode speed: up to 4,000 tokens a second, the company says.
  • Disputed claim: the launch claim of matching DeepSeek V4 Flash is contested, with one commenter reporting weaker tool selection than FunctionGemma on their own test.

Summary written by TrustList from the business's own published information, Nov 2026.

Benchmarks & AI stats

BenchmarkOfficialCommunity avg
No benchmark scores yet. Be the first to add one.

“Official” values are editor-approved and feed the ranking. “Community avg” is the mean of member submissions (shown for transparency; it never affects the ranking until an editor approves a value).

Sign in to add a benchmark score for this model.

Needle 3 launches

Trust Score

4/ 100
Trust Score: New

An earned signal from verification, reviews, awards, transparency and engagement - the vendor can't buy it.

Verification
0/100 · 20%
Reviews
0/100 · 30%
Awards
0/100 · 15%
Transparency
18/100 · 20%
Engagement
0/100 · 15%
Joining soon
Recommendations
Coming soon
Complaints
Coming soon

Updated 9/19/2026

Request a demo or quote from Needle 3

Protected by reCAPTCHA - Google Privacy Policy and Terms apply.

By sending, you agree we may share your request and contact details with the provider once you confirm your email.

Reviews

Write the first review of Needle 3

Used it? Your experience helps other buyers decide.

Write a review

Questions & answers

No questions yet. Be the first to ask about Needle 3.

Needle 3 alternatives

Similar listings buyers compare against this one.

Microsoft AI's text-to-speech model for keeping one voice across 23 languages, with voice cloning from short audio

Aleph Alpha's open-weight English-German mixture-of-experts model for self-hosting, released under Apache 2.0 in October 2026

Cantina Security's first open-weight security model, a post-trained GLM-5.3-Flash for vulnerability research and exploit checks

Open-source 2B decision model from AWS's Strands Agents project that picks options and returns scores with a reliability figure

Microsoft AI's real-time speech-to-text model for voice agents, transcribing 60 languages with continuous language detection

Anthropic's smallest and fastest Claude 5.5 model, available on its own API and three cloud platforms

JetBrains' open 12B mixture-of-experts coding model with 2.5B parameters active per token under Apache 2.0

Perplexity's open-weight decision model that picks among given options for routing and checks in AI systems