I Gave a Robot Duck a 15MB AI Brain (Needle 3)

Needle 3 from Cactus Compute is a tiny tool-calling model that can be cut down anywhere from 2 to 20 layers, so the same weights can fit almost any device. We fine-tuned it on an RTX 5090 to control a virtual Open Duck Mini robot that only obeys if you say please, gets angry when you're rude, and reacts to dramatic events like the floor being lava. On 49 hand-written prompts the model never saw in training, the fine-tuned version jumped from 27% to 84% accuracy, and even the 15MB 4-layer version more than doubled the original score. 🔗 Relevant Links Needle 3: https://cactuscompute.com/needle Open Duck Mini: https://github.com/apirrone/Open_Duck_Mini Needle Duck App: https://github.com/andrisgauracs/needle-duck ❤️ More about us Radically better observability stack: https://betterstack.com/ Written tutorials: https://betterstack.com/community/ Example projects: https://github.com/BetterStackHQ 📱 Socials Twitter: https://twitter.com/betterstackhq Instagram: https://www.instagram.com/betterstackhq/ TikTok: https://www.tiktok.com/@betterstack LinkedIn: https://www.linkedin.com/company/betterstack 📌 Chapters: 0:00 The polite robot duck 0:37 What is Cactus Compute? 1:04 Needle 2 recap 1:54 The intelligence ladder 2:23 Bigger model, same speed 2:44 Benchmarks and embeddings 3:09 Fine-tuning Needle 3 3:37 Meet the Open Duck Mini 4:04 Real physics in MuJoCo 4:48 Stock Needle 3 test 5:37 The training data 7:18 Fine-tuning on an RTX 5090 7:49 The fine-tuned duck 8:25 New prompts it has never seen 8:43 Where it still fails 9:05 Scoreboard results 9:58 The confidence score trade-off 10:28 Final thoughts