Cactus Needle 3: tiny on-device tool-calling models

Cactus Compute has released Needle 3, a family of extremely small language models built for on-device tool calling and automation. According to the model repository, the models range from just 8MB to 29MB — small enough to ship inside an app rather than call out to a server.

The pitch is that these models can handle the structured, function-calling work that powers automation flows: interpreting a request, choosing a tool, and formatting the call. Cactus claims Needle 3 can match DeepSeek V4 Flash on automation tasks, a notable assertion for models this size, though independent benchmarks aren't yet part of the record.

Why it matters

Most tool-calling today runs in the cloud, which adds latency, cost, and privacy tradeoffs. A model measured in megabytes changes the calculus:

• It can run locally on phones and edge hardware with no network round trip.
• Under-1B parameters keeps memory and battery demands low.
• The Apache 2.0 license lets developers embed and modify it freely.

The headline number to watch is real-world reliability: automation is unforgiving of malformed tool calls, and sub-30MB models leave little room for error. If Cactus's claims hold up under scrutiny, Needle 3 points toward a future where routine agentic tasks happen entirely on the device in your hand.