Small models and AI interface updates #18
Today's Letter
Needle, 26M tool-calling model released
- cactus-compute released Needle, a 26M-parameter function-calling model positioned for very small devices.
- The project says Needle was distilled from Gemini 3.1 into a Simple Attention Network architecture.
- The README says the model can be fine-tuned locally on a Mac or PC instead of requiring larger training hardware.
- Reported production performance on Cactus is 6000 toks/sec prefill and 1200 decode speed.
- Published model details include d=512, 8H/4KV, and a BPE vocabulary size of 8192.
- Weights are open under Cactus-Compute/needle, and the repository also says dataset generation materials are available.
Source: github.com
Google DeepMind outlines AI pointer principles
- Google DeepMind published a May 12, 2026 post describing an AI-era mouse pointer that can understand both on-screen context and user intent.
- The post framed the goal as moving AI out of a separate chat window and into existing workflows, so users do not need to drag content into a standalone tool.
- One example described pointing at an image of a building and asking for directions, with the system inferring the relevant context from the screen state.
- Google said the pointer has changed little in more than half a century and positioned the work as a redesign of a long-stable desktop interaction model.
- The materials referenced four guiding principles behind the interaction approach rather than a full product specification or API launch.
- The post focused on principles and demos, while also saying Google is starting to apply the approach in Chrome.
- Google also said Magic Pointer will roll out soon in Googlebook, but the post did not provide a full public launch timeline.
Source: deepmind.google
More: sources.news
Jocoletter curates AI, software, and product trends for developers and builders.
#GoogleDeepMind #cactus-compute
