# Needle: A 26M-Param Model Built for On-Device Tool Calls

> Cactus Compute's tiny encoder-decoder is distilled specifically for function calling at the edge, trading general chat for a narrow, useful job.

Published by The Open Weights on Mar 16, 2026. Canonical: https://theopenweights.com/news/needle-yfob

## Key facts

- Company: Cactus Compute
- Model: Needle
- Version: 1
- Category: Code
- Modalities: Code, Text / LLM
- License: MIT (Open weights, commercial use allowed)
- Parameters: 26M
- Architecture: NeedleForCausalLM
- Precision: BF16
- Weights size: 0.1 GB
- Significance: notable
- Published: 2026-03-16
- Last verified: 2026-07-23
- Hugging Face: https://huggingface.co/Cactus-Compute/needle
- Canonical URL: https://theopenweights.com/news/needle-yfob

Cactus Compute has released **Needle**, a compact encoder-decoder model with roughly 26 million parameters that is distilled for a single, practical task: tool and function calling. Rather than aiming to be a general-purpose chatbot, Needle is designed to interpret a request and emit the structured call needed to invoke an external function, according to the project's [GitHub repository](https://github.com/cactus-compute/needle).

The headline number here is size. At 26M parameters, Needle sits far below the sub-billion-parameter models that already market themselves as lightweight. That footprint is the point: it is small enough to run directly on phones, embedded hardware, and other resource-constrained devices without leaning on a cloud endpoint.

## Why it matters

Function calling is the connective tissue of modern AI agents — the step where a model decides which tool to invoke and with what arguments. Running that step locally has real advantages:

- **Latency**: on-device inference avoids a network round trip.
- **Privacy**: user prompts and parameters need not leave the device.
- **Cost**: no per-call API charges for a routine, high-frequency task.

By narrowing scope to tool calling and distilling aggressively, Needle bets that many agentic workflows don't need a large model to route requests — a specialized small one will do.

The model is released under the permissive **MIT license** and is available on [Hugging Face](https://huggingface.co/Cactus-Compute/needle). As with any distilled, task-specific model, the open questions are how reliably it parses complex or ambiguous calls and how well it generalizes across tool schemas it wasn't trained on. For developers building edge-first agents, though, a purpose-built model this small is an appealing starting point.

## Get the model

- [Hugging Face](https://huggingface.co/Cactus-Compute/needle)

## Sources

- [Cactus-Compute/needle](https://huggingface.co/Cactus-Compute/needle) — Hugging Face, Mar 16, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Needle: A 26M-Param Model Built for On-Device Tool Calls", The Open Weights, Mar 16, 2026, https://theopenweights.com/news/needle-yfob