December 18, 2025
Tiny bot, big drama
FunctionGemma 270M Model
Google's tiny offline AI lands; devs scramble for n8n hooks and cheer the December speedrun
TLDR: Google launched FunctionGemma, a small AI that runs offline and turns plain requests into app actions. The community split between hyped speed and real-world integration: some cheer the Ollama-ready drop and privacy perks, others debate n8n support and whether fine-tuning is a feature or a chore.
Google just dropped FunctionGemma, a pocket-sized AI that can actually do things—like setting reminders or toggling your phone’s flashlight—without the internet. It’s built to run on small devices and translate plain English into app actions. The community instantly turned this into a vibe check: a Google research lead, canyon289, popped in with “Happy to answer” energy, which gave the thread that “inside scoop” feel. Then the race began: nateb2022 posted the Ollama link, sparking a download dash, while SpaceManNabs lobbed the practical grenade—“can you run this from n8n?”—kicking off integration debates about whether this is plug-and-play or a fine-tuning project.
The fine-tuning twist (accuracy jumps from 58% to 85%) split the crowd: fans call it a “small specialist,” skeptics call it homework. Meanwhile, xnx fanned the hype with “Unbelievable shipping velocity” and a speedrun tweet, fueling jokes that Google is on a December launch binge. And orliesaurus’s “Im so dumb…” became the accidental meme template for devs facepalming through set-up snafus. Privacy diehards love the local-first promise; others wonder how “offline” it really is if tougher tasks still get bumped to bigger models. Drama level: high, download fingers: faster.
Key Points
- •Google released FunctionGemma, a function-calling–tuned variant of the Gemma 3 270M model for on-device agents.
- •FunctionGemma can operate offline as an independent agent or route complex tasks to Gemma 3 27B in compound systems.
- •Fine-tuning on "Mobile Actions" boosted reliability, increasing accuracy from 58% to 85%.
- •The model is engineered for edge devices (e.g., NVIDIA Jetson Nano, phones) and efficiently tokenizes JSON and multilingual inputs using a 256k vocabulary.
- •Broad ecosystem support spans fine-tuning (Transformers, Unsloth, Keras, NeMo) and deployment (LiteRT-LM, vLLM, MLX, Llama.cpp, Ollama, Vertex AI, LM Studio).