Today, I'm ranking the most popular

local AI use cases for you based on real

engineering experience. For each

category, I'll recommend the one model

or tool that I would start with so you

can skip hours of research and just get

going. Personally, I've spent hundreds

of hours testing with my RTX 1590, but a

lot of cases work on weaker hardware,

too. So, with that being said, let's get

right into it. First up, we're going to

start with local code autocomplete. And

this is straight up S tier. Code

autocomplete is one of the first ways

that we were using AI for automated

coding with something like ghetto

copilot. When you type code, the model

finishes your line or even fills in a

function body before you can think about

it. And while agent coding seems to be

taking over the world, code autocomplete

is still extremely useful. And the great

part is that it absolutely works great

with local models. One model you could

use is something like quen 2.5 coder

because it's a small 7 billion parameter

model which can run at sub 100

millisecond latency even sometimes on

GPUs which is a couple gigabytes of

VRAM. Now you're going to sometimes even

get results that are faster than

networkbased autocompletes because

there's no round trip. You're just you

know executing this with local models

and this gives you a similar experience

as the code autocomplete models that

were state-of-the-art a couple of years

ago. You can pair your local model with

something like continue dev and use for

the inference and then you've got a

self-hosted co-pilot replacement at the

very least you know the autocomplete

features of it that costs nothing after

hardware. While this doesn't beat the

newest agentic features that you get

with things like cloud code, copilot,

codecs, it's a great start and this

actually works well locally. Local can

beat cloud models here on the metric

that matters most when it comes to

autocomplete, which is response speed,

and you can fully customize it yourself

as well. We're going to be talking about

other coding use cases later, but first,

let's move to something different, photo

enhancement. This is something that I

would comfortably place in a tier. Photo

enhancement covers things like

upscaling, face restoration, background

removal, noise reduction, colorization,

anything basically where you take an

existing image and you make it better

without generating new content. I'll

focus on upscaling for now as this is an

example that is very common but the

whole category is strong locally. A tool

that you can get started with easily is

something like upscaly. It's a free open

source desktop app that uses real ESR

GAN models. You don't have to screw

around too much with Python. You don't

need the command line yet. No custom GPU

configuration just to see what's

possible. You drag a photo in and you

will get a 4x or even 8x upscaled

version out of it. From there, because

you can see what's possible, you can go

ahead and try and play around with the

models yourself and create a custom

workflow that works for whatever you

need to do. It's great to be able to

enhance your photos without having to

rely on only a cloud-based version of

Photoshop. And any dedicated GPU with

something like 4 GB of VRAM can handle

this use case in just seconds. Another

great use case is home automation. Home

automation covers many different tasks,

but the whole category can definitely be

a tier. With home automation, you can

have use cases like having your security

cameras detect a person in your driveway

and your phone getting a notification

with a snapshot, which will all be

processed on a box in your closet. And

the great part about this is that you

will not have a cloud subscription

anymore. And maybe more importantly,

your security footage leaving your

network to be stored on maybe some

random Chinese cloud that you don't know

about. Object detection is the headline

feature here, but local home automation

can also cover voice control, presence

sensing, and even energy monitoring. In

any case, the stack that you can start

with is Frigot NVR plus Home Assistant.

This is probably the most mature local

AI ecosystem that exists right now,

especially for beginners. With this, you

can have use cases like person

detection, vehicle detection, pets,

packages, license plates, and even some

basic facial recognition as part of some

of the latest Frigot versions. And all

of this will be processed entirely on

your hardware. Home Assistant has over 2

million active installations, and Frigot

has over 30,000 stars on GitHub. So,

you're not going to be alone. It's not

just a hobby project. It will be a full

ecosystem that you can plug into. The

only thing keeping this from S tier is

the set of complexity. You often will

have to configure Docker, camera

streams, detection zones, and even if

it's running, it might break in a couple

of weeks because things update fast in

the home automation space. Now, you

might be already thinking, "This is a

nice tier list so far, but how do I get

started?" Well, I put together a

collection of my own open- source

projects for many of the local AI use

cases we're covering today. The link is

in the description, and you should

definitely check it out. A whole

different use case is video generation,

and this one is a little bit

disappointing. I'm going to put it in

the C tier because the idea behind video

generation is that you can type a prompt

and get a full video clip out. Something

you can use for product demos, chosen

reader content, or just content

visualization. This is the category

everyone wants to work locally because

cloud video generation is expensive at

rate limited. Problem is though that

it's still very expensive to generate

locally in terms of the time that it

takes and the quality is just not great.

One model you can try is WAN 2.1 from

Alibaba, which can beat Sora on several

benchmarks, which was, you know, a

state-of-the-art model a while back. The

issue is though that even on my 1590, I

can't really use the full 14 billion

model. I already have to kind of go down

to the smaller 5 billion parameter one,

and the quality is just a lot less. In

addition, because it takes a long time

to generate video, it is also a lot more

time expensive to generate many variants

to figure out how you can generate a

video clip that actually looks good

compared to image generation, which

we'll get to later. You don't just

generate one frame. To get a good video,

you need to generate many different

frames, and that just takes a long time.

locally. This might be usable for some

experimental social media clips and

concept work, but real professional

production is still cloud territory, and

that's why this is going in C tier.

Speaking of something that's related,

image generation. Now, this works much

better. I would put image generation in

the S tier. With image generation, you

describe an image and the model can

create it. You can use it for

thumbnails, marketing assets, product

mockups, and concept art. This can also

include inpainting, outpainting,

basically editing. This is Disney image

and even generating a new image variant

based on an image that you're passing to

it. Now I'll focus on text to image

generation because that's where most

people start. The model you can use here

is Flux 2D which you can run through com

UI on something like my 5090. This is a

very comfortable use case and it's very

easy for me to generate an image in just

a couple of seconds. Unlike video

generation, this means that you can

generate hundreds of variants in, you

know, a short amount of time, which

increases the chances that you actually

get an image that you're happy with. And

in fact, Flux is a really good model

because in some blind tests, it achieved

a 71% win rate over some of the older

midjourney versions for editorial photo

realism. So, the models are really

getting much better. And with image

generation, I've also found that there's

a really great community where folks

have fine-tuned a lot of models for

custom characters and styles. And

there's even a lot of image generation

models that have less content filters

than the one in the cloud. It's good

that cloud has content filters,

especially for copyright, but sometimes

they go a little bit too far and they

become a bit unusable to be honest. And

actually training a custom Laura model

or actually fine-tuning one only takes

15 to 20 images. And you can do that on

consumer hardware as well. Now, one

thing that I've learned from using AI

image generation for my own content is

that iterative editing is where it gets

a bit tricky. Image properties like

gradients and complex textures make it

very difficult for those local models to

edit images without losing quality.

Personally, this is still where I use

the latest NA banana model or just use

Photoshop to just make sure that I can

use some of the cloud features. So, even

though it's S tier, there are some use

cases where it's going to fall behind a

little bit, but it is really getting

better every single month. A completely

different but fun use case is voice

agents. Now, this one is something that

I'm going to place in C tier because of

the wide variety of use cases which can

work in a variety of ways. Basically,

the idea behind voice agents is that you

can talk to your computer and it will

talk back like a local Alexa or customer

support bot that runs on your hardware.

But ideally, it should also be able to,

you know, do actions for you. For

example, you can use it to do hands-free

home automation. The best open source

option right now would be something like

pipecat which can chain together a

speechto text model an lm and then a

texttospech model into one single

pipeline and that will allow you to get

something like sub 800 millisecond voice

to voice latency on some standard

hardware on something like Mac OS. The

problem is that the model size

constraint means that the AS responses

locally are noticeably dumber than you

can get from the best local models that

you interact with over chat and of

course the best cloud models. In fact,

this is even an issue with cloud voice

agents. They can achieve 200 to 500

millisecond latency, but their

intelligence is more around GBT40 level

and not the latest models that are out

there. And of course, you know, this

intelligence gap is much higher for

local voice agents. That being said, if

you just have a singular command that

you want to run, voice agents can work

pretty well. The main issue where they

lack quality is in longer conversations

because with local models, you're just

much more constrained on your context

window, which means that as you have a

long conversation with a voice agent, it

will simply go off track and you don't

have that issue if your voice agent is

just there to process one command and

execute it. So that would be the

recommended use case I would say. Now,

one component of voice agents is a text

to speech model and you can use text to

speech on its own. And on its own, this

is definitely a tier. With text to

speech, you feed text in and you get

natural sounding audio out. You can use

this for audiobook narration, voice over

for video, accessibility features in

your app, or even trying to clone your

own voice for content, which I've tried,

but that one is a bit ambitious. It

doesn't work very well. Now, this

category also corporates music

generation and sound effects because

it's not really my specialty. In general

though, voice synthesis is where local

models have made the biggest gains. Now,

text to speech has had the most dramatic

transformation in my opinion of any

local AI category in the past 18 months

because I would say that it's almost a

solved problem at this point, especially

for the English language. The model that

I would start with is Chatterbox from

resemble AI because it beats 11 laps in

a couple of blind tests with over 60%

listener preference rates. And the base

model might be English only, but you can

even go to chatterbox multilingual which

covers 23 plus languages. And so the gap

versus a cloud offering like 11 labs has

closed for a lot of use cases. And in

some cases the gap is gone entirely.

There are some gotchas. A lot of models

will hallucinate or really degrade in

quality past about a thousand

characters. But there's a lot of ways to

solve that. You know, you can split up

your text in multiple batches and just

process them one by one. Now, if you're

finding this tier list useful, make sure

to hit subscribe because over 90% of you

are missing out on the latest AI

engineering news that's based on reality

and not hype. So, make sure to join the

club. Next, we're going to be talking

about speech to text, which is another

component of voice agents that you can

extract on its own, and it's another

solved problem in my opinion. Speech to

text is absolutely S tier. You record

audio or you pass an audio file, you can

get a text transcript back for meeting

notes, podcast transcription, subtitle

generation, or just turning your voice

memos into searchable text. I use this

constantly myself for processing my own

YouTube video content. And for English,

it works very well. A model you can use

is faster whisper with large V3 Turbo,

which gives you over four times speed

over the original Whisper model. And the

real workflow that I use is a two-stage

pipeline. I use whisper to get a fast

accurate raw transcription and then I

use local LLM to clean up filler words

and extract the core meaning and this

way I can store the notes into something

like my Obsidian vault to reference

later. The transcription itself is

almost instant but the LM cleanup step

does add some time. The thing is though

I can just run it in the background so

that's not an issue at all for me. The

gap of cloud models is mainly with

speaker diorization, which means that

you can split up some audio context into

the different speakers. So let's say you

have a meeting with 10 people, you want

to be able to know who's speaking which

sentences, right? Cloud models are a

little bit better at doing that, but I

fully believe that local models will

catch up very quickly. The core use case

of transcribing audio just works very

well nowadays with local models. Next,

we're going to be talking about OCR,

optical character recognition, which I'm

going to comfortably place as our first

B tier. OCR covers table extraction,

formula recognition, and just converting

scan documents into structured data,

which is useful for so many use cases.

One tool to start with is Sura, or it

can use a deepse OCR model that has been

released recently. Now, a lot of the use

cases are a little bit more on the

boring side, like being able to transfer

and process invoices, but it does work

pretty well. Next, we're going to be

talking about agentic coding, which is

much more interesting than just doing

code autocomplete and seems to be hyped

up more and more for local models. Now,

the one disappointing thing about agent

coding is that it doesn't work that well

unless you have very good hardware. A

lot of YouTube videos, they show Agenta

coding and then they show how it can

generate Python code and that's good and

all, but with Agenta coding, you really

want a model to be competent enough to

read your entire codebase, write code,

run tests, and iterate until your

feature works, the feature that you

asked for with text. The issue is though

that most absolutely most hardware is

too incompetent for genting. Now, I've

created many videos on my channel

covering how performant agent coding is

on my 5090, and it's definitely getting

closer, but it is nowhere near as useful

as using code autocomplete, nor

something like an AI chat, because

eventually local AI agents choke on

larger projects. That being said, this

is something that improves every single

month. So definitely check out the

videos that I have in my channel which

I'll put in the link in the description

below to get the most out of your local

setup because it's very difficult to set

up properly and you really need a bit of

an expert guide to do this right with

aentic coding you might be able to match

something like GBT40 locally but Claude

Opus 4.6 and similar models have really

up the game here and there is just a

huge gap between what you can do locally

versus with the state-of-the-art cloud

models. Even if your local model is

intelligent enough, the problem is that

it will get slow as the context window

fills up. And unlike with these other

use cases where you can clear out the

context window after you improve small

bits of the task, with a coding, you

cannot clear the context window with

every step. You need the coding model to

have a good idea of how your codebase

works. So in the end, these models are

going to work very slowly on your local

hardware, especially compared to cloud

models, unless you have a very beefy PC.

Most people promoting local models with

Agenta coding are not using it for

serious projects. And if you don't

believe me, check out some of my master

classes on this channel to learn the

truth. If you do want to get started

with a coding, I recommend some of the

latest Quen models. But again, I cover

that extensively in the videos on my

channel already, so go check those out.

As models get stronger, I expect it to

move to A or S tier, especially as local

models will catch up as well. You know

what isn't A tier yet, though? AI chats.

A traditional use case. With AI chats,

you can ask questions, brainstorm ideas,

summarize documents, draft emails, the

same stuff you would use something like

chat GPT for, but it can run entirely on

your machine with no data leaving your

network. I would recommend new models

like something like Quen 330B through LM

Studio. And this is a pretty optimized

model that can run pretty fast, but

honestly, there's many choices for AI

chat models. You can use a newer Mistral

model or even check out the older but

still competent Open- source OpenAI

model that's out there. In any case,

locally you can really mimic what GPT40

was able to do a year plus ago and now

fully locally. You can use it for so

many different purposes that it's just a

great use case altogether and a very

beginner friendly one because everyone

understands the concept of an AI chat,

right? The great part too is that you

have many choices for the right user

interface here. You can use something

like LM Studio which is a desktop app

that already has a chat UI or even

create your own web app and customize it

to your liking. There are so many open

source repos and again check out the

link in the description if you want to

make a good start on that because I've

got many templates for you to customize.

A specific subset of AI chats is rack

retrieval augmented generation which is

a very fancy word that some of you

probably already know about which is the

fact that you can point an AI at your

own files, company documents, research

papers or notes and it will answer

questions based on that content. It's

pretty similar to AI chat but the thing

is is that generally it's a bit more

expensive and complex to set up. You

might have to put your documents into a

vector database and you have to make

sure that those documents are retrieved

properly to make sure that the AI chats

can answer questions about them. But if

you do it right, well, this can work

very well for your own use cases and to

make sure that these local models are up

to date with the latest knowledge in

your use case, which is not in the

training data of the model, especially

because some of these open source models

have been trained half a year or a year

ago. So, rag is a very important

paradigm and I'm putting it in beats

here because of the technical complexity

to set it up, but it is kind of a

requirement for a lot of real AI chat

use cases. So, it's definitely something

you want to learn to set up yourself.

One tool you can start with is something

like open web UI, which gives you a full

rag pipeline out of the box with a clean

interface. And the nice part about this

is that because it's all running

locally, data privacy is the consistent

top reason enterprises will use these

kinds of self-hosted LLMs. So if you

learn these skills, you can definitely

get an AI engineering job with that as

well. So we have talked about some AI

use cases now like a chat, a rack chat,

we have agentic coding, code

autocomplete. But what about the ability

to create any AI agent of your dream

where I'm defining an agent as a system

that can autonomously make decisions,

execute actual actions for you, and be

able to solve a problem in many

different ways. Well, the thing is a

true AI agent is very difficult to run

locally. I'm going to put it in Ctier,

but I wanted to make sure I explained

this to you because you've probably seen

many videos here on YouTube talking

about AI agents. The problem is though

that most of those videos are not real

agents. They're more deterministic

workflows on something like any 10 with

a small LM component in the middle that

might make one or two decisions. But a

true AI agent that can run like cloud

code simply requires a very good

language model or else it will get

confused about all the tools it has

access to. it will not be able to run

autonomously without you pushing it into

the right direction and you will just

find that there are many issues with it

in general. That being said, I'm not

saying that AI agents are incapable to

work locally. It just depends on your

use case and most people that I've

talked to who want to build their first

AI agent are a little bit too ambitious.

They might want to build a research

agent that works better than something

you can use with OpenAI GPT Pro. But to

be quite honest, it's very difficult to

build a better AI agent than using a

platform that you can access over the

cloud. Again, something like cloud code

because even with cloud code, you can

point it to a local model, but it just

won't perform in the same autonomous

way. That being said, AI agents in

particular change all the time, and I

expect that as models get better, this

will move up to the B tier and the A

tier eventually. But even then, AI

agents simply won't run on weak hardware

because of literal mathematical

constraints that I've covered in other

videos on my channel. So depending on

your use case, this is just not going to

work very well. But if you think that

that's not true and you have had good

experiences creating local AI agents, I

would love to hear it from you in the

comments down below. But please tell me

what use case you have because most

people trying to run local AI agents

aren't really running through agents.

They're just running regular workflows

with LLMs sprinkled somewhere in

between. But how about a more simple AI

assistant that you can run locally like,

you know, OpenClaw, an always on

personal AI that manages your calendar,

tries your email, summarizes your day,

and handles whatever you throw at it for

personal use cases. The local version of

Siri or Google Assistant, but running

private and 24/7 on your own hardware.

So tools like OpenClaw aim to be exactly

this and sound great on paper. The issue

is though that setting them up properly

in a secure way that makes sure that you

actually keep your account secure is

pretty difficult. It does require a

little bit of security knowledge.

Because of that complexity, I would

right now put AI assistance into the

beats here because it can work quite

well if you are very aware of the

security problems with something like

OpenClaw. Now, personally, I would still

use OpenClaw with a state-of-the-art

cloud model because they're much better

protected against things like

jailbreaks. But there may be one

exception where I would use local

models, which is for well-defined cron

jobs, which are things that run on a

schedule like every 8 hours. This might

be something where you are summarizing

your feed into a daily digest or just

classifying your incoming emails. For

something like this, a local 14 billion

model works just fine. And yes, you can

use a local model with that. So far, we

have a lot of use cases and they all are

pretty competent, right? Even the ones

in Ctier, they're very usable and

they're getting better every single day.

But then what is the Dtier for? Well, to

be honest, that's for vibe coding. Vibe

coding with local models just doesn't

work. The idea with vibe coding is that

you just describe an app in plain

English and the AI will build the whole

thing and you never read the code. You

just judge whether the result works. And

this is different from agentic coding

because you are not explicitly reviewing

or steering anything. Now, honestly,

this workflow needs a frontier model to

cover for the fact that you're not

reviewing at all what it writes. A small

language model can code quite well with

the right guidance in a couple of files.

And for models under 14 billion

parameters, you cannot even use tool

calling properly, which is a requirement

for proper agenda coding, let alone vibe

coding where you're not guiding the

model at all. Vive coding with cloud

models already has serious problems.

Researchers found security

vulnerabilities in one out of 10 lovable

generated apps, for example. But with

weaker local models, those problems

multiply. So, let's recap this tier list

a little bit. The three S tier use cases

generally match or sometimes even beat

cloud models. Code autocomplete, image

generation, and speech to text. The

pattern here is that some of the more

boring use cases consistently outperform

the hyped ones for local models.

However, as models get better, some of

the more complex workflows like AI

agents as well as voice agents will just

get better over time and hopefully

everything will be in the B tier or

above in a couple of years from now. And

if you want to get started with local AI

projects, you should check out the link

in the description below and get started