Needle 3
Needle 3 is a 121M model that selects declared tools and fills their arguments from evidence in a user request. It compiles tool schemas into a constrained decoding grammar and withholds calls when required values are missing or unsupported.
What it does
Developers can use it to run tool calls, extract structured data, or generate embeddings locally on supported devices.
- Availability
- Try it in the browser; Python package and source are on GitHub
- Pricing model
- Not established
- Deployment
- On-device; platform engines and air-gapped setup are documented
- Privacy
- All three jobs run on the device.
- Licence
- Not established
Capability
Modality
Access
Deployment
Pricing
Target user
Evidence
8 claims- Each argument is derived from a span of the request, and a required argument without a span or default withholds the call. [1]
Each argument is copied from a span of the request, and the model writes that derivation as its reasoning line before the call.
Cactus Compute. Quoted text is outside CC BY 4.0 and keeps its original rights. - Ranges, enums, patterns, lengths, and item counts are compiled into the decode grammar. [1]
Ranges, enums, patterns, lengths and item counts are compiled into the decode grammar, so an invalid value is unrepresentable rather than merely discouraged.
Cactus Compute. Quoted text is outside CC BY 4.0 and keeps its original rights. - Above five tools, retrieval selects the five closest tools for the request. [1]
Above that, retrieval engages: every schema is embedded once, each request is embedded, and only the five closest tools enter the context with the grammar rebuilt over them.
Cactus Compute. Quoted text is outside CC BY 4.0 and keeps its original rights. - Needle 3 supports tool calls, structured extraction, and text embeddings. [2]
- **Tool calls**: given the functions your app exposes, Needle picks the right ones and fills every argument from what the user said.
Cactus Compute. Quoted text is outside CC BY 4.0 and keeps its original rights. - The model uses a byte-level grammar compiled from schemas to constrain output. [2]
The weights are compressed to CQ2-bit with Cactus Quants; a byte-level grammar compiled from your schemas constrains every token, and every response carries a calibrated confidence score from a learned head.
Cactus Compute. Quoted text is outside CC BY 4.0 and keeps its original rights. - The repo includes the 20-layer needle3.cact model and a needle3.safetensors checkpoint to fine-tune. [2]
The repo holds the 20-layer `needle3.cact`, the `needle3.safetensors` checkpoint to fine-tune, and an engine per platform.
Cactus Compute. Quoted text is outside CC BY 4.0 and keeps its original rights. - Needle 3 runs on-device, with an engine provided for each platform folder. [2]
- The model can be tried in a browser, and the Python package can be installed with pip install cactus-needle. [2]
Sources
2
1
Cactus ComputeOwner-controlled
https://cactuscompute.com/blog/designing-tools-for-needle
Retrieved 2026-10-27 01:17 UTC
Verified
2
Cactus ComputeOwner-controlled
https://huggingface.co/Cactus-Compute/needle3
Retrieved 2026-10-27 01:17 UTC
Verified
Versions
6| Version | Effective | Change | Changed fields | State |
|---|---|---|---|---|
| 6 | 2026-10-18 15:45 UTC | evidence update | score | Current |
| 5 | 2026-10-16 17:30 UTC | evidence update | score | Current |
| 4 | 2026-10-06 03:19 UTC | material change | availability, deployment | Current |
| 3 | 2026-10-02 07:03 UTC | material change | capability, availability, deployment | Current |
| 2 | 2026-09-25 13:08 UTC | material change | capability, access, availability, deployment, privacy | Current |
| 1 | 2026-09-24 13:40 UTC | publication | None | Current |
Original catalog content is licensed under CC BY 4.0, attributed to wowsuchai. Source quotations and linked content keep their original rights.
