Google engineer explained how to fine-tune a tiny LLM from 46% to 90% accuracy on your phone in 21 minutes - better than $1500 on-device AI bootcamps. pick Gemma 270M -> generate synthetic task data -> fine-tune with LoRA -> quantize to int4 -> deploy to Pixel and hit... 2000 tokens per second. That loop is how a 270M model beats a 70B one on your task, running fully offline in your pocket. Gemma 270M + synthetic data + LoRA + int4 quantization + on-device runtime - that's the stack. Watch and save it, then fine-tune your own tiny agent tonight.show more

Here's the YouTube video with appropriate audio

Can you please do @aiDotEngineer a favor and not steal their content to promote your articles?

Kind of misleading opening line, you don't fine tune the model on your "phone", you just run it after its fine tuned elsewhere ✌️

This is actually super interesting. I'm going to watch! Cheers

Such a time waste just prompt this to fable 5

Why can't @Apple just use this model on the iPhone's

Great thoughts. See this does the orchestration for free

for a narrow set of tasks like this does, you could probably 100x this again with a 1.2M model from scratch.

its cool i did this a while back to turn it into a reasoning model. unfortunately, the reasoning didn't seem to do it much good and it was a little state machine. right when it was obvious, wrong when it wasn't

determinism is almost as fast as code

Another BS post

Is it too much to ask that devs learn to talk into the mic?

46 to 90 percent accuracy in 21 minutes on a phone is wild

This was painful to watch. Go on a presentation skills course. Learn to speak in full sentences, and stop losing focus like there was a fucking squirrel around every corner. I just couldn't get to the end. My loss, but think of the time I saved not shouting at the screen

I don't believe such a combo, especially on gemma, can make anything of quality

That article still treats each agent as a monolithic expertise container. Theyre just wrappers around the same LLM, so the real multi angle analysis comes from combining those LLMs in parallel.

Such a marketing post

@grok what's the filesize of Gemma 270M?

270M干翻70B，这不就是中国互联网的普世真理：别看脸大，能干活的才是真大佬。今晚就微调一个帮我自动回复甲方的小模型。

AI Engineer videos are a banger!

has anyone ever got gemma 270m to do anything ??

Tiny task-specific models win on efficiency; frontier agents win when work needs broad judgment. The useful comparison is cost per completed outcome:

What happens when 90% accuracy becomes a release gate for another task? Fine-tuning gains need a judge tied to deployment pass conditions. Practical gate:

@grok please summarize this and give me links on where I can learn more.

Small models are becoming much more interesting when they are trained for specific tasks.

2000 tokens/sec on a phone is wild, never hit anything close locally

That article still treats each agent as a monolithic expertise container. Theyre just wrappers around the same LLM, so the real multi angle analysis comes from combining those LLMs in parallel.

If you only have mac m3 16gb ram, this is not for you 😂

Note100 : 99.5

smt new thx for sharing

Google is a government surveillance program.

HOOLYYY HOLLLYYYYYY

small models keep getting more capable

Great topic. I cover that in my new book -