You can run a frontier-grade AI on your laptop with no monthly bill and no internet.

@keshavsuki

3.3K followers · usually 602 views a post

Views

10.3K

Likes

479

Comments

50

Shares

167

Selling · Course / Info

Claude Code

https://keshavsuki.com ↗

Why it worked

The video provides high-value, technical 'how-to' content that solves a specific pain point (cost and privacy of AI) while positioning the creator as an expert, making the CTA feel like a natural extension of the tutorial.

Summary

The creator explains how to use model distillation to create a small, efficient AI model that runs locally on a laptop. He outlines a three-step process involving Distilabel, Unsloth, and Ollama to achieve high performance without internet or API costs.

Structure

1Hook: Run frontier AI locally for free
2Explain the concept of model distillation
3Step 1: Use Distilabel to generate training data
4Step 2: Use Unsloth to fine-tune the model
5Step 3: Run the model locally using Ollama
6CTA: Comment to get the build

Product placement

Claude Code (mentioned in title and bio link), Distilabel, Unsloth, Together AI, Ollama

Original caption

I cloned a 70B model into a 3B model that runs free and offline on my MacBook. Total cost to build: $12 #ClaudeCode #AIAutomation #llm #anthropic #aiagents

More from @keshavsuki

Their other posts in the index, biggest breakout first.

Call to action

Comment "clone" and I'll send you the full build.

On-screen text

Claude Code You Can Run Frontier AI Offline On Your Laptop For Free GIANT MODEL (70B) cleaning apparatus CLEANED MODEL (3B) Keeps Smarts What is knowledge distillation? Knowledge distillation is a machine learning technique that aims to transfer the learnings of a large pre-trained model, the "teacher model", to a smaller "student model". It's used in deep learning as a form of model compression and knowledge transfer, particularly for massive deep neural networks. PULLING STRUCTURED DATA FROM MEDICAL NOTES INPUT: UNSTRUCTURED MEDICAL NOTES EXTRACTION ENGINE DATA/ distilabel Synthesize data for AI and add feedback on the fly! STRUCTURED TRAINING DATA PIPELINE TRAINING EXAMPLE A DATA INPUT 1 TRAINING EXAMPLE B DATA INPUT 2 NEW Introducing Unsloth Studio Run and train AI models locally with Unsloth Studio. Export Model Choose where to save your exported model. Save Locally Push to Hub Base Model Checkpoint Export Method Cancel Start Export

Transcript

You can run a frontier-grade AI on your laptop with no monthly bill and no internet. And before you tell me these small local models are too dumb for real work, you are completely correct. So you don't download a small one. You take a giant one and clean it down into a tiny one that keeps the smarts. Here's exactly how. A small model is dumb on its own, but you can teach it using a big one. It's a term called distillation and it costs about $12. Here's the system. Step one, pick your task. Say pulling structure data out of medical notes. A platform called Distilabel calls a 70 billion teacher model onto Together AI and has it generate thousands of perfectly labeled examples. Then uses the same teacher as a judge to keep only the best. But there's a catch. Distilabel validates the whole pipeline before it even spends a cent. So run the dry run first and make sure there's an end token between your training examples, or the model will leak context. Step two, a platform called Unsloth fine tunes a 3 billion model on those examples in about 90 minutes on one GPU. Then exports it as a small compressed file. Step three, you drop that file into Ollama and watch it run locally. Now the 3 billion model runs 60 tokens a second on a MacBook, parsing a medical note into clean JSON in under a second. No internet, no API key, nothing leaving the machine. Comment "clone" and I'll send you the full build.