A 14MB agent model that runs on a $200 phone
Cactus Compute put Needle 2 on Show HN: 45 million parameters, compressed into a single 14MB binary, running a full session in 28MB of RAM. Apache 2.0, weights on Hugging Face, runner on GitHub.
It doesn't chat. It doesn't write prose. It does tool calling, device control and structured extraction, which is a deliberate and slightly brutal scoping decision — strip out everything an agent's dispatcher never needs and see how small the dispatcher gets. The answer is 14MB. It scores 63.7% on Google's Mobile Actions benchmark, decodes 500+ tokens per second on a Raspberry Pi 5, 400 to 1,500 on a Quest 3S, and 300 to 700 on sub-$200 Android phones. No GPU, no NPU. It'll run on an ESP32-S3.
The comparison they lead with is that it trades wins with FunctionGemma 270M, LFM2.5 230M and Apple's on-device foundation model while being 5x to 70x smaller. Trading wins with things an order of magnitude bigger is the claim worth checking, and the weights are right there to check it with.
Same day as Meta's 30B local agent model, which is a nice accident. Glimmer is the agent that thinks on your laptop. Needle 2 is the agent that lives in your light switch. The interesting question is how much of an agent's actual work is the second kind — parse the intent, pick the tool, fill the arguments — and whether that part was ever worth 30 billion parameters.
https://cactuscompute.com/needle
← Back to all articles
It doesn't chat. It doesn't write prose. It does tool calling, device control and structured extraction, which is a deliberate and slightly brutal scoping decision — strip out everything an agent's dispatcher never needs and see how small the dispatcher gets. The answer is 14MB. It scores 63.7% on Google's Mobile Actions benchmark, decodes 500+ tokens per second on a Raspberry Pi 5, 400 to 1,500 on a Quest 3S, and 300 to 700 on sub-$200 Android phones. No GPU, no NPU. It'll run on an ESP32-S3.
The comparison they lead with is that it trades wins with FunctionGemma 270M, LFM2.5 230M and Apple's on-device foundation model while being 5x to 70x smaller. Trading wins with things an order of magnitude bigger is the claim worth checking, and the weights are right there to check it with.
Same day as Meta's 30B local agent model, which is a nice accident. Glimmer is the agent that thinks on your laptop. Needle 2 is the agent that lives in your light switch. The interesting question is how much of an agent's actual work is the second kind — parse the intent, pick the tool, fill the arguments — and whether that part was ever worth 30 billion parameters.
https://cactuscompute.com/needle
Comments