Needle — a 26M-parameter LLM for IoT & edge devices 🔧 • Runs on microdevices and embedded hardware • Supports function calling • Offline-first design

Needle scales down Gemini to 26M parameters. Devs get a streamlined, potent model without the usual bloat. Easier integration, same capability. Who's ready to upgrade?! #OpenSource #AI

A 26-million parameter model just matched Gemini's tool-calling skill. That's not magic. It's distillation. And it's one of the most important tricks in modern AI.

Tiny models keep getting unfairly good. Needle just distilled Gemini tool calling into 26M parameters by literally deleting the FFN from the encoder blocks. Just attention and a gated residual…

My favorite detail about the new Needle 26M tool calling model is the architecture. They completely dropped the feed-forward networks from the encoder. It just runs pure self-attention, then decodes…

Evening Learnings 🌆 5 AI concepts worth understanding today.

Great to see forward thinking small models like this!

is top of HN today. He shrunk a frontier model into something small enough to fit on a wristwatch, kept only the function-calling, and trained it in 27 hours 🔥

tool calling in a 26m model is more interesting than another giant chatbot.

Cactus 把 Gemini 3.1 拆了——只拿走工具调用,蒸馏成 26M 开源模型 Needle。 选工具比大它十倍的 FunctionGemma 还准。端上 6000 tok/s prefill,工具决策几十毫秒。 下一代 agent 栈,可能比 26M 微模型 dispatch。 #AI #LocalLLM