Needle 3
A 121M-parameter open model that does one job on tiny devices: turn a plain request into tool calls or structured data.
In short. Needle 3 does not chat. You give it a list of functions your app offers, such as set the lights or start a timer, and a request from the user. It picks the right function and fills in the arguments. It can also pull typed fields out of messy text. It is small enough to run on a phone, a watch or a Raspberry Pi.
Summary
Needle 3 is a tool-calling model from Cactus Compute, released in mid-September 2026 under Apache 2.0. It is not a general chatbot. Given the functions an app exposes and a user request, it returns one JSON object with the calls to make and the arguments for each, plus a confidence score. Asked for two things, it returns two calls in order. Asked for something no tool covers, it returns an empty list instead of guessing. Needle 3 can also extract structured data: you declare a record, hand over text, and a decoding grammar guarantees the output parses as that shape.
The architecture is unusual. Cactus calls it a laddered Simple Attention Network, with 20 layers, a width of 768 and an 8,192-token context. Instead of feed-forward layers it uses Hadamard-based channel mixing and hashed n-gram lookup tables called engrams, and the company says about 70.8 million of the 121 million parameters live in those tables. The training data, 360 billion tokens of structured examples, is proprietary. Every depth from 2 to 20 layers is trained as its own model, so the same weights can be sliced into a deployable subnetwork of 29 million parameters or more. The Hugging Face repository holds a 242 MB half-precision checkpoint and a 35.3 MB .cact file with weights stored at about 2 bits each for the Cactus engine.
The developer says Needle 3 beats models ten times its size on mobile tool-call benchmarks such as Mobile Actions, DroidCall and BFCL v4, and matches models two to three times bigger on extraction tests such as DSTC8 and SNIPS. It also says that, after fine-tuning on DroidCall, subnetworks of 4 layers and up pass DeepSeek V4 Flash on that test. On a Raspberry Pi 5 it quotes 400 to 4,000 tokens per second for decoding. These comparisons come from the developer's charts, were run against baselines it chose, and we have not reproduced them.
The limits are stated plainly in the developer's own notes. The model was trained with small tool sets, and accuracy falls as the list grows past roughly two dozen tools, so large catalogs should be shortlisted first. This release ships without a contrastive embedding head, so its embedding call returns a pooled signal from the confidence head and is not a trained retrieval embedding. The engine is under 1 MB per platform and prebuilt for more than a dozen targets, installed with pip install cactus-compute, and the company documents LoRA fine-tuning on a single machine.
Good for
- Voice and text control of phones, watches, smart-home devices, robots and cars, where a tiny offline model is enough
- Turning notifications, invoices or messages into typed fields without sending data to a cloud service
- A fast first step in a hybrid setup, using the confidence score to decide when to act, confirm or hand over to a larger model
- Commercial products, since the license is Apache 2.0
Watch out for
- It is a specialist and cannot hold a conversation or write long answers
- Accuracy falls when you give it more than about two dozen tools at once, so shortlist large catalogs
- The embedding call is not a trained retrieval embedding in this release
- The context window is 8,192 tokens
- Benchmark claims come from the developer's own charts and have not been independently checked
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Needle 3 (20 layers, FP16 checkpoint) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Needle 3 (20 layers, FP16 checkpoint) FP16 | 121M | 0.278 GB | 0.139 GB | 0.0835 GB | Laptop or phone |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Needle 3?
- Needle 3 does not chat. You give it a list of functions your app offers, such as set the lights or start a timer, and a request from the user. It picks the right function and fills in the arguments. It can also pull typed fields out of messy text. It is small enough to run on a phone, a watch or a Raspberry Pi.
- Can I use Needle 3 commercially?
- Yes. Needle 3 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Needle 3 open source?
- Needle 3 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Needle 3 on my own computer?
- Yes. The smallest variant, Needle 3 (20 layers, FP16 checkpoint), needs about 0.0835 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Needle 3 released?
- Needle 3 was released by Cactus Compute on September 2026.
- How long a context window does Needle 3 have?
- Needle 3 supports a context window of up to 8K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Cactus Compute: Needle 3 cactuscompute.com
- Cactus Compute: Intelligence ladders (17 Sep 2026) cactuscompute.com
- Cactus Compute: Porting Needle 3 (18 Sep 2026) cactuscompute.com
- Needle 3 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Whistle Cactus Compute | 55M | – | Apache 2.0 (Permissive) | 0.253 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Ternary Bonsai 2 PrismML | 2.21B – 27.4B | 256K | Apache 2.0 (Permissive) | 0.473 GB Laptop or phone |