Needle Models by Cactus Compute
Last refreshed 2026-09-17. Next refresh: weekly.
Details
Capabilities
About
Needle is Cactus Compute's family of automation / tool-calling foundation models for tiny devices. Line spans original Needle (~26M, 2026-05-12), Needle 2 (45M / 14MB CQ2-bit, 2026-08-09), and tip Needle 3 (sliceable 2–20 layer ladder, 25–121M / 8–29MB CQ2-bit, 2026-09-17). Built on Simple Attention Networks; oriented to tool calling, structured extraction, and on-device embeddings rather than open chat. Tip priority this pass: Needle 3 with prior Needle 1 + Needle 2 siblings under the same family. Compare it for Open Source / Agents / on-device / tool-use.
Decision facts
- Best fit
- agentson deviceJSON / Tool use
- Capability starting point
- Needle 3 with 8k context and JSON / Tool use and structured outputs
- Lowest tracked input
- Not tracked
- Closest related family
- K2 Horizon
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs on device, 8k context, and 121M parameters.
Use when the workload needs on device, 2k context, and 45M parameters.
Use when the workload needs on device, 26M parameters, and JSON / Tool use.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Needle 3 | Use when the workload needs on device, 8k context, and 121M parameters. | 2026-09 | on device8k context121M parameters | Current |
| Needle 2 | Use when the workload needs on device, 2k context, and 45M parameters. | 2026-08 | on device2k context45M parameters | Current |
| Needle | Use when the workload needs on device, 26M parameters, and JSON / Tool use. | 2026-05 | on device26M parametersJSON / Tool use | Current |