Needle2: a 14MB Apache-2.0 agent model runs on a $200 phone
A 45M-param tool-caller that runs offline in 28MB of RAM on a Pi or budget phone — plus Upstage's Solar Pro 4 lands as a cheap agentic API.
Copy markdown
14MB, 45M params, and it fits on a microcontroller
Cactus's Needle2 squeezes a 45M-parameter tool-calling model into 14MB with 2-bit quantization, running a session in about 28MB of RAM. It clocks 500+ tokens/sec on a Raspberry Pi 5 and 300-700 on sub-$200 phones — no GPU, no cloud.
Apache 2.0 weights, already in a shipping wearable
The weights are on Hugging Face under Apache 2.0 with the runtime on GitHub, and Pebble already ships Needle2 in its Index Ring. You can fine-tune it on your own laptop in minutes to hours to teach it a custom tool vocabulary.
Build this weekend: a fully offline device agent
Needle2 handles tool-calling, device control (smart home, wearables, robots, apps), and schema-based extraction entirely on-device. That's a voice-controlled agent for a Quest 3S, an ESP32-S3, or a $200 Android that keeps working with zero network.
It trades wins with models 5-70x its size
On the 961-row Mobile Actions benchmark Needle2 hits 63.7% overall and 98.3% on name accuracy, holding its own against FunctionGemma 270M and LFM2.5 230M despite being a fraction of their size.
Upstage ships Solar Pro 4, a cheap agentic API
Released today, Solar Pro 4 scores 57 on Terminal-Bench v2.1 (+13.8 over Solar Pro 3), carries a 512K context, and costs $0.30/$1.20 per M tokens — 90% off through Sep 10 on Upstage Console and OpenRouter. It's API-only, not open weights.