Cactus Needle 3: Lightweight automation models matching larger LLM performance.

What it does

Cactus Needle 3 is a compact foundation model designed for device-local inference. The model runs as a single binary ranging from 8 to 29 MB and performs three core functions: tool calls, structured extraction, and text embeddings. When given a set of available functions, it routes user requests to the correct ones and fills in arguments automatically. For extraction tasks, developers declare a data shape and the model returns parsed, typed fields. The same weights also generate embeddings for semantic search and matching without network calls.

Who it is for

The model targets developers building features for resource-constrained environments: mobile phones, wearables, smart home devices, robots, automotive systems, microcontrollers, and AR glasses. The pitch emphasizes on-device processing—reducing latency, protecting privacy, and avoiding cloud dependencies. Use cases include voice control for smart homes, natural-language instructions for robots, local phone assistants, and embedded search over device documents.

Pricing

Fine-tuning costs $19 per run, with hosted compute and training-data generation included. The site does not show prices for other services or usage tiers.

How it stands out

Needle trades general chat ability for narrow performance on automation tasks. The model is purpose-built for tool calling and structured extraction rather than conversational AI. A distinctive feature is "intelligence laddering"—the same weights can run at 2, 3, or up to 20 layers, creating a scalability spectrum from 8 MB to 29 MB binaries. The site claims fine-tuned 4-layer versions match larger models like DeepSeek V4 Flash on downstream tasks. The model also handles request-to-tool routing defensively: if no declared tool fits, it returns an empty list rather than guessing.

What a founder should check

First, verify whether on-device inference actually eliminates the competitive moat of cloud-hosted LLMs for the target verticals. Incumbents like Anthropic, OpenAI, and Gemini already run local models on phones; a founder should test whether Needle's tool-calling performance materially outpaces theirs on real mobile hardware. Second, examine switching costs and the fine-tuning requirement. The $19 fine-tuning fee is low, but deeply integrated tool-calling systems may lock customers in through data and task-specific training rather than price. Third, assess pricing pressure from open-source alternatives and larger models shipping local inference by default—whether the market will tolerate a paid service around a model function that increasingly ships free with devices.

Thinking of building something like this?

Every launch here is a competitor to somebody's idea. If yours is close, check it against the market before you build: the Full Check names the rivals, the prices and the gaps.