You're tired of paying so much for an AI.

@hatsunama

3.2K followers · usually 1.1K views a post

Views

15.3K

Likes

812

Comments

113

Shares

167

SoundTikTok sound · original sound - hatsunama

Promoted product

Unsloth Studiounsloth.ai

Unsloth is an open-source, no-code web UI for training, running and exporting open models in one unified local interface.

Why it worked

The video provides high-value, actionable technical information that solves a specific pain point (high cost of AI) while positioning the creator as an expert, making the mention of the tool feel like a helpful resource rather than a sales pitch.

Summary

The creator provides a curated list of open-source AI models categorized by hardware requirements, ranging from mobile phones to enterprise-grade data centers. He highlights specific models like Bonsai 27B, Gemma 4 12B, and Qwen3.6, while mentioning Unsloth Studio as a platform for running and training these models.

Structure

1Introduction to local AI models
2Phone-tier model recommendation
3Laptop-tier model recommendation
4

Original caption

There is an Ai for every phone or laptop. Here is where you can find yours. Hope this helps! #ai #news #llm #localai #fyopppppppppppppppppppp

More from @hatsunama

Product placement

Unsloth Studio is mentioned as a platform to run and train models.

Call to action

Hit that save button.

On-screen text

LOCAL AI FOR ALL HARDWARE! PrismML Announces 1-bit Bonsai 27B - The First 27B Model to Run on a Phone Gemma 4 12B Qwen3.6 - How to Run Locally antirez/deepseek-v4-gguf How much VRAM does GLM-5.2 (744B-A40B MoE) need? nvidia/GLM-5.2-NVFP4

Transcript

You're tired of paying so much for an AI. That's why you're here looking for a local one. Well, here's the full list from phone to data center of all the best open source AIs that you can find right now. Hit that save button. Let's get in. First, for your phone, it's about a 4 gigabyte download and it is Bonsai 27B. This is the first 27 billion parameter model that could actually run on your phone. Multimodal, agentic tools, 256k character context window, 90% of the full quality you get from regular top of the line AIs. Next is the nano tier and this is a 12 gigabyte download will run on most all laptops. Of course, I'm talking about Gemma 4 12B. This is Google's efficient multimodal model and it will work like I said on most all laptops. Next is our micro tier, a 36 gigabyte download. Of course, this is Qwen3.6 the 27B model and this one absolutely crushes it on all the benchmarks and on several places punches way above its belt. For this you will need an RTX 4090 at the very least. Next is what we're calling the small version, this is 128 gigabytes. This is Deepseek 4 flash and it is a huge 248 billion parameter model. For this you're really gonna need a high RAM workstation. This is out of reach probably for most people honestly. Then we've got the medium tier which is 256 gigabyte download. Of course I'm talking about GLM 5.2 the massive 744 billion parameter model. You technically still could run this at home if you had the maxed 256 gigabyte Mac Studio, you could still run this in your house. And then our largest model just for fun is a 512 gigabyte download. This is GLM 5.2-NVFP4 and this of course is enterprise grade performance where you'll need like 8 Blackwell GPUs to run it.