YT YouTube2Text

YouTube2Text → Transcripts → The Ultimate Local AI Tier List For 2026

The Ultimate Local AI Tier List For 2026 — Transcript

by Zen van Riel · 4,689 words · 687 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00Today, I'm ranking the most popular
0:01local AI use cases for you based on real
0:03engineering experience. For each
0:05category, I'll recommend the one model
0:07or tool that I would start with, so you
0:09can skip hours of research and just get
0:11going. Personally, I've spent hundreds
0:13of hours testing with my RTX 4090, but a
0:16lot of cases work on weaker hardware,
0:18too. So, with that being said, let's get
0:20right into it. First stop, we're going
0:22to start with local code auto complete,
0:24and this is straight up S-tier.
0:27Code auto complete is one of the first
0:28ways that we were using AI for automated
0:31coding with something like GitHub
0:32Copilot. When you type code, the model
0:34finishes your line or even fills in a
0:36function body before you can think about
0:38it. And while agentic coding seems to be
0:40taking over the world, code auto
0:41complete is still extremely useful. And
0:44the great part is that it absolutely
0:45works great with local models. One model
0:48you could use is something like Qwen 2.5
0:50Coder, because it's a small 7 billion
0:53parameter model, which can run at
0:55sub-100 millisecond latency, even
0:57sometimes on GPUs with just a couple
0:59gigabytes of VRAM. Now, you're going to
1:02sometimes even get results that are
1:03faster than network-based auto
1:04completes, because there's no round
1:06trip. You're just, you know, executing
1:08this with local models. And this gives
1:11you a similar experience as the code
1:12auto complete models that were
1:14state-of-the-art a couple of years ago.
1:16You can pair your local model with
1:18something like Continue Dev and use
1:20Ollama for the inference, and then
1:21you've got a self-hosted Copilot
1:24replacement, at the very least, you
1:25know, the auto complete features of it,
1:27that costs nothing after hardware. While
1:29this doesn't beat the newest agentic
1:31features that you get with things like
1:33Cloud Code, Copilot, Codex, it's a great
1:35start, and this actually works well
1:37locally. Local can beat cloud models
1:40here on the metric that matters most
1:42when it comes to auto complete, which is
1:43response speed, and you can fully
1:45customize it yourself, as well. We're
1:46going to be talking about other coding
1:48use cases later, but first, let's move
1:49to something different: photo
1:51enhancement. This is something that I
1:53would comfortably place in A-tier. Photo
1:55enhancement covers things like
1:57upscaling, face restoration, background
1:59removal, noise reduction, colorization,
2:02anything basically where you take an
2:03existing image and you make it better
2:05without generating new content. I'll
2:08focus on upscaling for now as this is an
2:09example that is very common, but the
2:11whole category is strong locally. A tool
2:14that you can get started with easily is
2:15something like Upscaley. It's a free,
2:18open-source desktop app that uses
2:19Real-ESRGAN models. You don't have to
2:22screw around too much with Python, you
2:24don't need the command line, yet no
2:25custom GPU configuration. Just to see
2:27what's possible, you drag a photo in and
2:30you will get a 4x or even 8x upscaled
2:32version out of it. From there, because
2:35you can see what's possible, you can go
2:36ahead and try and play around with the
2:38models yourself and create a custom
2:40workflow that works for whatever you
2:41need to do. It's great to be able to
2:43enhance your photos without having to
2:45rely on only a cloud-based version of
2:47Photoshop. And any dedicated GPU with
2:50something like 4GB of free RAM can
2:52handle this use case in just seconds.
2:55Another great use case is home
2:57automation. Home automation covers many
3:00different tasks, but the whole category
3:02can definitely be A-tier. With home
3:05automation, you can have use cases like
3:06having your security cameras detect a
3:08person in your driveway and your phone
3:10getting a notification with a snapshot,
3:12which will all be processed on a box in
3:14your closet. And the great part about
3:16this is that you will not have a cloud
3:18subscription anymore and maybe more
3:20importantly, your security footage
3:22leaving your network to be stored on
3:24maybe some random Chinese cloud that you
3:26don't know about. Object detection is
3:28the headline feature here, but local
3:30home automation can also cover voice
3:32control, presence sensing, and even
3:34energy monitoring. In any case, the
3:36stack that you can start with is Frigate
3:38NVR plus Home Assistant. This is
3:40probably the most mature local AI
3:42ecosystem that exists right now,
3:44especially for beginners.
3:46With this, you can have use cases like
3:48person detection, vehicle detection,
3:50pets, packages, license plates, and even
3:52some basic facial recognition as part of
3:55some of the latest Frigate versions. And
3:57all of this will be processed entirely
3:59on your hardware. Home Assistant has
4:01over 2 million active installations, and
4:03Frigate has over 30,000 stars on GitHub,
4:05so you're not going to be alone. It's
4:06not just a hobby project. It will be a
4:08full ecosystem that you can plug into.
4:10The only thing keeping us from S tier is
4:12the setup complexity. You often will
4:14have to configure Docker, camera
4:15streams, detection zones, and even if
4:17it's running, it might break in a couple
4:19of weeks because things update fast in
4:21the home automation space. Now, you
4:23might be already thinking, "This is a
4:25nice tier list so far, but how do I get
4:27started?" Well, I put together a
4:29collection of my own open-source
4:30projects for many of the local AI use
4:32cases we're covering today. The link is
4:34in the description, and you should
4:35definitely check it out. A whole
4:36different use case is video generation,
4:39and this one is a little bit
4:40disappointing. I'm going to put it in
4:41the C tier because the idea behind video
4:44generation is that you can type a prompt
4:45and get a full video clip out. Something
4:48you can use for product demos,
4:49children's video content, or just
4:51content visualization. This is the
4:52category everyone wants to work locally
4:54because cloud video generation is
4:56expensive and rate-limited. Problem is
4:58though that it's still very expensive to
5:00generate locally in terms of the time
5:01that it takes, and the quality is just
5:03not great. One model you can try is When
5:052.1 from Alibaba, which can beat Sora on
5:08several benchmarks, which was, you know,
5:10a state-of-the-art model a while back.
5:13The issue is though that even on my
5:141590, I can't really use a full 14
5:17billion model. I already have to kind of
5:19go down to the smaller 5 billion
5:21parameter one, and the quality is just a
5:24lot less. In addition, because it takes
5:26a long time to generate video, it is
5:28also a lot more time expensive to
5:30generate many variants to figure out how
5:32we can generate a video clip that
5:34actually looks good. Compared to image
5:36generation, which we'll get to later,
5:37you don't just generate one frame. To
5:39get a good video, you need to generate
5:41many different frames, and that just
5:42takes a long time. Locally, this might
5:46be usable for some experimental social
5:48media clips and concept work, but real
5:50professional production is still cloud
5:52territory, and that's why this is going
5:54in C tier.
5:55Speaking of something that's related,
5:57image generation, now this works much
5:59better. I would put image generation in
6:01the S tier. With image generation, you
6:03describe an image and the model can
6:05create it. You can use it for
6:06thumbnails, marketing assets, product
6:08mock-ups, and concept art. This can also
6:11include in-painting, out-painting,
6:13basically editing existing image, and
6:15even generating a new image variant
6:17based on an image that you're passing to
6:19it. Now, I'll focus on text-to-image
6:21generation because that's where most
6:23people start. The model you can use here
6:25is Flux 2 Dev, which you can run through
6:27ComfyUI. On something like my 5090, this
6:30is a very comfortable use case, and it's
6:32very easy for me to generate an image in
6:34just a couple of seconds. Unlike video
6:36generation, this means that you can
6:37generate hundreds of variants in, you
6:39know, a short amount of time, which
6:41increases the chances that you actually
6:43get an image that you're happy with.
6:45And in fact, Flux is a really good model
6:47because in some blind tests, it achieved
6:49a 71% win rate over some of the older
6:52Midjourney versions for editorial
6:54photorealism. So, the models are really
6:56getting much better.
6:59And with image generation, I've also
7:00found that there's a really great
7:01community where folks have fine-tuned a
7:04lot of models for custom characters and
7:06styles, and there's even a lot of image
7:08generation models that have less content
7:10filters than the one in the cloud. It's
7:12good that cloud has content filters,
7:14especially for copyright, but sometimes
7:16they go a little bit too far, and they
7:17become a bit unusable, to be honest. And
7:19actually training a custom LoRA model,
7:22or actually fine-tuning one, only takes
7:2450 to 20 images, and you can do that on
7:26consumer hardware, as well. Now, one
7:28thing that I've learned from using AI
7:30image generation for my own content is
7:31that iterative editing is where it gets
7:33a bit tricky.
7:35Image properties like gradients and
7:36complex textures make it very difficult
7:38for those local models to edit images
7:41without losing quality. Personally, this
7:43is still where I use the latest native
7:45been added model or just use Photoshop
7:47to just make sure that I can use some of
7:49the cloud features. So, even though it's
7:50S tier, there are some use cases where
7:53it's going to fall behind a little bit,
7:54but it is really getting better every
7:56single month. A completely different but
7:58fun use case is voice agents. Now, this
8:00one is something that I'm going to place
8:01in C tier because of the wide variety of
8:04use cases where it can work in a variety
8:07of ways. Basically, the idea behind
8:08voice agents is that you can talk to
8:10your computer and it will talk back like
8:12a local Alexa or customer support bots
8:13that runs on your hardware. But ideally,
8:15it should also be able to, you know, do
8:17actions for you. For example, you can
8:19use it to do hands-free home automation.
8:22The best open source option right now
8:23would be something like Pipe Cat, which
8:25can chain together a speech to text
8:27model, an LLM, and then a text to speech
8:29model into one single pipeline. And that
8:32will allow you to get something like sub
8:34800 millisecond voice to voice latency
8:36on some standard hardware on something
8:39like macOS.
8:41The problem is that the model size
8:43constraints means that the AI's
8:44responses locally are noticeably dumber
8:47than you can get from the best local
8:48models that you interact with over chat,
8:50and of course, the best cloud models. In
8:53fact, this is even an issue with cloud
8:54voice agents. They can achieve 200 to
8:56500 millisecond latency, but their
8:58intelligence is more around GPT-4 level
9:01and not to the latest models that are
9:02out there. And of course, you know, this
9:05intelligence gap is much higher for
9:07local voice agents. That being said, if
9:10you just have a singular command that
9:11you want to run, voice agents can work
9:13pretty well. The main issue where they
9:15lack quality is in longer conversations
9:17because with local models, you're just
9:19much more constrained on your context
9:21window, which means that as you have a
9:22long conversation with a voice agent, it
9:25will simply go off track. And you don't
9:27have that issue if your voice agent is
9:29just there to process one command and
9:31execute it. So, that would be the
9:33recommended use case, I would say. Now,
9:35one component of voice agents is a text
9:37to speech model and you can use text to
9:39speech on its own and on its own this is
9:41definitely a tier. With text to speech
9:43you feed text in and you get natural
9:45sounding audio out. You can use this for
9:47audiobook narration, voiceover for
9:48video, accessibility features in your
9:50app, or even trying to clone your own
9:52voice for content which I've tried but
9:54that one is a bit ambitious it doesn't
9:55work very well. Now this category also
9:57covers music generation and sound
9:59effects because it's not really my
10:01specialty. In general though voice
10:03synthesis is where local models have
10:04made the biggest gains. Now text to
10:06speech has had the most dramatic
10:08transformation in my opinion of any
10:10local AI category in the past 18 months
10:12because I would say that it's almost a
10:14solved problem at this point especially
10:16for the English language. The model that
10:18I would start with is Chatterbox from
10:20Resemble AI because it beats 11 labs in
10:23a couple of blind tests with over 60%
10:26listener preference rates. And the base
10:28model might be English only but you can
10:30even go to Chatterbox multilingual which
10:32covers 23 plus languages. And so the gap
10:35versus a cloud offering like 11 labs has
10:37closed for a lot of use cases and in
10:40some cases the gap has gone entirely.
10:43There are some gotchas. A lot of models
10:45will hallucinate or really degrade in
10:46quality past about 1,000 characters but
10:50there's a lot of ways to solve that. You
10:51know, you can split up your text in
10:53multiple batches and just process them
10:55one by one. Now if you're finding this
10:56tier list useful make sure to hit
10:58subscribe because over 90% of you are
11:00missing out on the latest AI engineering
11:02news that's based on reality and not
11:05hype so make sure to join the club. Next
11:07we're going to be talking about speech
11:08to text which is another component of
11:10voice agents that you can extract on its
11:12own and it's another solved problem in
11:14my opinion. Speech to text is absolutely
11:17S tier.
11:18You record audio you pass an audio file
11:20you can get a text transcript back. For
11:23meeting notes, podcast transcription,
11:24subtitle generation, or just turning
11:27your voice memos into searchable text. I
11:30use this constantly myself for
11:31processing my own YouTube video content
11:34and for English it works very well. A
11:36model you can use is faster Whisper with
11:38large V3 Turbo which gives you over four
11:41times speed over the original Whisper
11:42model. And the real workflow that I use
11:45is a two-stage pipeline. I use Whisper
11:47to get a fast, accurate raw
11:49transcription and then I use local LLM
11:52to clean up filler words and extract the
11:54core meaning. And this way I can store
11:56the notes into something like my
11:57Obsidian vault to reference later. The
11:59transcription itself is almost instant,
12:02but the LM cleanup step does add some
12:04time. The thing is though, I can just
12:05run it in the background, so that's not
12:07an issue at all for me.
12:09The gap of cloud models is mainly with
12:10speaker diarization, which means that
12:13you can split up some audio context into
12:16the different speakers. So let's say you
12:17have a meeting with 10 people, you want
12:19to be able to know who's speaking which
12:21sentences, right? Cloud models are a
12:23little bit better at doing that, but I
12:25fully believe that local models will
12:26catch up very quickly. The core use case
12:28of transcribing audio just works very
12:30well nowadays with local models.
12:33Next we're going to be talking about
12:34OCR, optical character recognition,
12:36which I'm going to comfortably place as
12:38our first beat here.
12:40OCR covers table extraction, formula
12:42recognition, and just converting scanned
12:43documents into structured data, which is
12:45useful for so many use cases. One tool
12:48to start with is Suriya, or you can use
12:50a Deep Seek OCR model that has been
12:51released recently. Now, a lot of the use
12:53cases are a little bit more on the
12:54boring side, like being able to transfer
12:56and process invoices, but it does work
12:59pretty well. Next we're going to be
13:00talking about agentic coding, which is
13:02much more interesting than just doing
13:03code autocomplete and seems to be hyped
13:05up more and more for local models. Now,
13:08the one disappointing thing about
13:09agentic coding is that it doesn't work
13:11that well unless you have very good
13:13hardware. A lot of YouTube videos they
13:15show agentic coding and then they show
13:17how it can generate Python code and
13:18that's good and all, but with agentic
13:20coding you really want a model to be
13:22competent enough to read your entire
13:24codebase, write code, run tests, and
13:26iterate until your feature works, the
13:28feature that you asked for with text.
13:31The issue is though that most absolutely
13:33most hardware is too incompetent for
13:35agentic coding.
13:37Now, I've created many videos on my
13:38channel covering how performant agentic
13:41coding is on my 5090, and it's
13:43definitely getting closer. But, it is
13:45nowhere near as useful as using code
13:47autocomplete, nor something like an AI
13:50chat, because eventually local AI agents
13:53choke on larger projects. That being
13:55said, this is something that improves
13:56every single month. So, definitely check
13:59out the videos that I have in my
14:00channel, which I'll put in the link in
14:01the description below, to get the most
14:03out of your local setup, because it's
14:04very difficult to set up properly, and
14:06you really need a bit of an expert guide
14:08to do this right. With agentic coding,
14:10you might be able to match something
14:11like GPT-4o locally, but Claude Opus 4.6
14:14and similar models have really upped the
14:16game here, and there is just a huge gap
14:18between what you can do locally versus
14:19with the state-of-the-art Claude models.
14:22Even if your local model is intelligent
14:24enough, the problem is is that it will
14:26get slow as the context window fills up.
14:28And unlike with these other use cases,
14:29where you can clear out the context
14:31window after you improve small bits of
14:34the task, with agentic coding, you
14:36cannot clear the context window with
14:37every step. You need the coding model to
14:40have a good idea of how your code base
14:41works. So, in the end, these models are
14:44going to work very slowly on your local
14:46hardware, especially compared to cloud
14:48models, unless you have a very beefy PC.
14:50Most people promoting local models with
14:52agentic coding are not using it for
14:54serious projects. And if you don't
14:55believe me, check out some of our
14:57masterclasses on this channel to learn
14:59the truth. If you do want to get started
15:01with agentic coding, I recommend some of
15:02the latest Qwen models, but again, I
15:04covered that extensively in the videos
15:06on my channel already, so go check those
15:08out. As models get stronger, I expect it
15:10to move into A or S tier, especially as
15:13local models will catch up as well. You
15:15know what isn't A tier yet, though? AI
15:17chats, a traditional use case. With AI
15:19chats, you can ask questions, brainstorm
15:21ideas, summarize documents, draft
15:23emails, the same stuff you would use
15:25something like ChatGPT for, but it can
15:27run entirely on your machine with no
15:29data leaving your network. I would
15:30recommend new models like something like
15:32Coin 330B through LM Studio. And this is
15:36a pretty optimized model that can run
15:37pretty fast, but honestly, there's many
15:39choices for AI chat models. You can use
15:41a newer Mistral model or even check out
15:44the older, but still competent
15:46open-source OpenAI model that's out
15:47there. In any case, locally, you can
15:49really mimic what GPT-4 was able to do a
15:52year plus ago and now fully locally. You
15:55can use it for so many different
15:56purposes that it's just a great use case
15:58altogether and a very beginner-friendly
16:00one because everyone understands the
16:01concept of an AI chat, right? The great
16:03part, too, is that you have many choices
16:05for the right user interface here. You
16:07can use something like LM Studio, which
16:09is a desktop app that already has a chat
16:10UI, or even create your own web app and
16:12customize it to your liking. There are
16:14so many open-source repos, and again,
16:16check out the link in the description if
16:18you want to make a good start on that
16:19because I've got many templates for you
16:21to customize.
16:22A specific subset of AI chats is RAG,
16:25retrieval augmented generation, which is
16:26a very fancy word that some of you
16:28probably already know about, which is
16:29the fact that you can point an AI at
16:31your own files, company documents,
16:33research papers, or notes, and it will
16:35answer questions based on that content.
16:37It's pretty similar to AI chats, but the
16:39thing is is that generally it's a bit
16:40more expensive and complex to set up.
16:42You might have to put your documents
16:44into a vector database, and you have to
16:46make sure that those documents are
16:48retrieved properly to make sure that the
16:50AI chats can answer questions about
16:52them. But if you do it right, well, this
16:54can work very well for your own use
16:56cases and to make sure that these local
16:58models are up to date with the latest
17:00knowledge in your use case, which is not
17:02in the training data of the model,
17:04especially because some of these
17:05open-source models have been trained
17:06half a year or a year ago. So, RAG is a
17:09very important paradigm, and I'm putting
17:11it in beats here because of the
17:12technical complexity to set it up, but
17:14it is kind of have
17:15for a lot of real AI chat use cases. So,
17:18it's definitely something you want to
17:19learn to set up yourself. One tool you
17:21can start with is something like Open
17:23Web UI, which gives you a full rack
17:25pipeline out of the box with a clean
17:26interface. And the nice part about this
17:29is that because it's all running
17:30locally, data privacy is the consistent
17:32top reason enterprises will use these
17:35kinds of self-hosted LLMs. So, if you
17:37learn these skills, you can definitely
17:38get an AI engineering job with that as
17:40well. So, we have talked about some AI
17:42use cases now like a chat, a rack chat,
17:45we have agent decoding, code auto
17:46complete, but what about the ability to
17:48create any AI agent of your dream? Where
17:51I'm defining an agent as assistant that
17:53can autonomously make decisions, execute
17:55actual actions for you, and be able to
17:58solve a problem in many different ways.
18:00Well, the thing is a true AI agent is
18:02very difficult to run locally. I'm going
18:04to put it in C tier, but I want to make
18:06sure I explain this to you because
18:08you've probably seen many videos here on
18:09YouTube talking about AI agents. The
18:11problem is though that most of these
18:12videos are not real agents. They're more
18:14deterministic workflows than something
18:16like Any Ten with a small LLM component
18:19in the middle that might make one or two
18:20decisions. But a true AI agent that can
18:23run like Claude Code simply requires a
18:25very good language model, or else it
18:28will get confused about all the tools it
18:30has access to. It will not be able to
18:32run autonomously without you pushing it
18:34into the right direction, and you will
18:35just find that there are many issues
18:37with it in general. That being said, I'm
18:39not saying that AI agents are incapable
18:41to work locally, it just depends on your
18:43use case. And most people that I've
18:45talked to want to build the first AI
18:47agents are a little bit too ambitious.
18:49They might want to build a research
18:50agent that works better than something
18:52you can use with OpenAI GPT Pro. But to
18:55be quite honest, it's very difficult to
18:57build a better AI agent than using a
18:59platform that you can access over the
19:01cloud, again something like Claude Code.
19:03Because even with Claude Code, you can
19:05point it to a local model, but it just
19:07won't perform in the same autonomous
19:09way. That being said, AI agents in
19:11particular change all the time, and I
19:13expect that as models get better, this
19:15will move up to the B tier and the A
19:17tier eventually. But even then, AI
19:20agents simply won't run on weak hardware
19:23because of literal mathematical
19:25constraints that I've covered in other
19:26videos on my channel. So, depending on
19:28your use case, this is just not going to
19:30work very well. But if you think that
19:33that's not true, and you have had good
19:34experiences creating local AI agents, I
19:37would love to hear it from you in the
19:38comments down below. But please tell me
19:40what use case you have because most
19:42people trying to run local AI agents
19:44aren't really running through agents.
19:46They're just running regular workflows
19:47with LLMs sprinkled somewhere in
19:49between. But how about a more simple AI
19:51assistant that you can run locally?
19:53Like, you know, Open Claw. An always-on
19:56personal AI that manages your calendar,
19:58triages your email, summarizes your day,
20:00and handles whatever you throw at it for
20:02personal use cases. The local version of
20:04Siri or Google Assistant, but running
20:06private and 24/7 on your own hardware.
20:09So, tools like Open Claw aim to be
20:11exactly this, and sound great on paper.
20:13The issue is though that setting them up
20:15properly in a secure way that makes sure
20:17that you actually keep your account
20:18secure is pretty difficult. It does
20:20require a little bit of security
20:22knowledge. Because of that complexity, I
20:24would right now put AI assistants into
20:25the B tier because they can work quite
20:27well if you are very aware of the
20:29security problems with something like
20:31Open Claw. Now, personally, I would
20:33still use Open Claw with a stated yard
20:34cloud model because they're much better
20:36protected against things like
20:37jailbreaks. But there may be one
20:39exception where I would use local
20:41models, which is for well-defined cron
20:43jobs, which are things that run on a
20:45schedule, like every 8 hours. This might
20:47be something where you are summarizing
20:49your feed into a daily digest or just
20:51classifying your incoming emails. For
20:53something like this, a local 14 billion
20:55model works just fine. And yes, you can
20:57use a local model with that. So far, we
20:59have a lot of use cases, and they all
21:01are pretty competent, right? Even the
21:03ones in C tier, they're very usable, and
21:05they're getting better every single day.
21:07But then, what is the D tier for? Well,
21:09to be honest, that's for vibe coding.
21:11Vibe coding with local models just
21:13doesn't work. The idea with vibe coding
21:15is that you just describe an app in
21:17plain English and the AI will build the
21:18whole thing and you never read the code,
21:21you just judge whether the result works.
21:23And this is different from agenda coding
21:24because you are not explicitly reviewing
21:26or steering anything. Now, honestly,
21:28this workflow needs a frontier model to
21:30cover for the fact that you're not
21:31reviewing at all what it writes. A small
21:34language model can code quite well with
21:36the right guidance in a couple of files.
21:38And for models under 14 billion
21:40parameters, you cannot even use tool
21:42calling properly, which is a requirement
21:44for proper agenda coding, let alone vibe
21:47coding where you're not guiding the
21:48model at all. Vibe coding with cloud
21:50models already has serious problems.
21:52Researchers found security
21:53vulnerabilities in one out of 10 lovable
21:56generated apps, for example. But with
21:58weaker local models, those problems
22:00multiply. So, let's recap this tier list
22:03a little bit. The three S tier use cases
22:06generally match or sometimes even beat
22:08cloud models. Code auto complete, image
22:10generation, speech to text. The pattern
22:12here is that some of the more boring use
22:14cases consistently outperform the hyped
22:16ones for local models. However, as
22:19models get better, some of the more
22:20complex workflows, like AI agents as
22:23well as voice agents, will just get
22:25better over time and hopefully
22:27everything will be in the B tier or
22:29above in a couple of years from now. And
22:32if you want to get started with local
22:33app projects, you should check out the
22:35link in the description below and get
22:37started today.

About this transcript

This page contains the full transcript of The Ultimate Local AI Tier List For 2026 by Zen van Riel, generated from the public captions YouTube serves with the video. The transcript has 4,689 words across 687 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.