In this quick video, we're going to go

over LM Studio Bionic, a new tool

released by LM Studio, which is

essentially a genetic harness for

open-source models, for open weights

models. And in my opinion, this is more

important than ever with everything

going on in AI. I do believe open-source

and open weights is something that

everybody should become familiar with.

LM Studio is one of my favorite ways to

run open-source models on my computer.

I've done a few videos on them already.

But LM Studio is pretty much just a

model runner, a way to run LLMs. And

today, it's not really about LLMs

anymore, it's about running LLMs with an

agents. So, ChatGPT has now ChatGPT work

and Codex, and that is essentially what

LM Studio is doing. But, the promise of

LM Studio is you could run it with local

models, open weight models, where you're

technically not paying a third-party

provider. Everything is technically

staying local on your computer, it's

private, and you have all these models

you can choose from that are optimized

to work on your computer. Now, I'm just

going to give you a first look. I've

been running this for last day since it

came out. I'm running it with Gemma 4

12B and GPT-OSS. I'm going to tell you

what I've discovered, the pros and cons.

But, first, let's just look at the

announcement really quickly. Bionic is

an AI agent for getting real work done

with open models, including coding,

research, and complex work with

documents and files. You can use local

models or switch to open-source models

in the cloud for heavier tasks, all

while staying in control of your privacy

and AI spend. For all LM Studio Bionic

users, we commit to zero data retention

and never training on your data. A

Bionic agent excels at coding and

document work, voice input with

state-of-the-art local voice

transcription, flexible model execution.

You can run it either locally, LM Link,

so LM Studio running on a different

computer, or use a largest frontier

open-source models through LM Studio

secure cloud. So, essentially, this is a

soft launch for their own cloud model

services, kind of like what Ollama has

been doing. They're also releasing this

offline voice transcription, kind of

like Whisper Flow or Aqua Voice, and is

running off of Vox Row by Mistral. And

you can apparently run it in almost any

app, so this may be a side feature that

might replace one of the subscriptions

you already have. It essentially has two

modes, just like ChatGPT and Claude,

work and code. For coding, you obviously

want it to have shell access, just like

Claude Code, Code X, etc. And here's the

thing, Bionic works with powerful

open-source models like GLM 5.2, Kimiko

2.7 code. I'm sure Kimiko 3 is going to

come soon. And by the way, we get a

trust me, bro, these Chinese models are

running on US infrastructure, but

they're not really going to too much

details, they're just saying trust me.

And for document work, slides and

sheets, it's more in a sandbox, just

like Claude Co-work. You can have it

organized local folders, edit files,

summarize materials, all running locally

on your computer with those open models.

They support checkpoints and rollbacks.

Here you can see the local models, and

what's cool about this is if you're

already running LM Studio on your

computer and you download a bunch of

local models, they obviously work now

within Bionic. The one thing is you have

to turn off your LM Studio server, which

is nice because you don't have to have

all these things running. That's how I

would connect other things to LM Studio

before. And then here they talk about

their cloud inference with zero data

retention with GLM 5.2, Kimiko 2, and

they point out here when using cloud

models, your requests are processed

transiently and are not retained after

the request completes. So, the

difference between this and previous LM

Studio is LM Studio was able to run LLMs

and use tools, but not make edits, not

really take a big actions. But Bionic is

able to do that. You create multiple

projects, we'll create a new one now.

We'll call this for video, and you can

see down here it could either be code or

work. So, just like Claude Co-work or

Claude Code. Now, by default, it starts

in work, and if you do that, it won't

have shell access. So, in my case, I

like to use code, and you can choose the

models. So, here we have my three models

I have installed at the moment. One

little bug I realized is that even

though we just created this new project,

you have to go and select it. At that

point, it will ask you what directory

you want to work in. So, we're just

going to say Bionic test, and we'll say

what tools you have access to. But this

release makes it kind of clear of what's

really going on in AI right now. The

main workflows are now work or code.

Frontier models are getting too

expensive, and people want to run and

probably will need to run open-source in

the near future. But again, for LM

Studio, I think the main release here is

their cloud inference, models that they

are providing with a subscription that

are hosted in the US. By the way, I

think that Ollama's cloud models aren't

officially guaranteed to run in the US,

but they mainly are. Whereas here

they're saying inference is being

provided on US servers. Again, trust me,

bro. So anyways, going back, we see here

it tells me the tools it has access to.

Now, the biggest caveat here or what to

take away from this is that open source

or open weights models that you're

running locally on your computer will

not compare with cloud code or codex or

anti-gravity or any of these frontier

models. These are trillion parameter

models that are running on huge

infrastructure that you just won't have

on your computer. That means you have to

adjust your approach. The approach in

how much context you want to use, the

approach in how much you're going to

trust your agent. Because even last week

there were plenty of reports of GPT 5.6

deleting files off of people's computer.

And that's not a dig at OpenAI. My point

is these agents still make mistakes. And

if we're talking about a small model, a

12 billion parameter model, or a 9

billion parameter model, or even a huge

model, they can still screw up. And when

you're giving these models access to

your computer, you have to understand

this is a smaller model. It has less

instructions, has less reasoning

capabilities, and it still has access to

your full file system. So keep all of

that in mind. That being said, if we go

into settings, we see here the ability

to log in. This is where you make your

LM Studio account and then access their

cloud models. I haven't done that yet.

You can turn on web search, but you need

to be logged in for that. In

integrations, they have connected apps,

MCB servers. So they have some out of

the box. And what I want you to focus on

here is unlike with LM Studio, which

ports over your models, it doesn't port

over your MCB servers. So you're going

to have to reinstall each MCB server. It

will support both local and remote MCB

servers. But what I want you to notice

here is one thing that's missing, and

it's crucial, and I'm sure it's coming,

is skills. They don't have support for

skills right now. And in coding agents

specifically, skills are so important

because they're reusable prompts. And

I've done plenty of videos on skills.

But the fact that skills are missing

here means that the CLI tools I'm using

are less powerful or less capable

because the Bionic agent can't read the

skills that go with the CLI tools. Now,

let's say you don't have LM Studio

installed and you want to just start

with Bionic and see all the different

models that are available, you could go

down to settings explore and choose

whatever model you want. And then the

most important setting that you should

probably change is model defaults. What

is the context window you want your

models to run? Now, this depends on your

computer. You have to calculate and you

can have Cloud Code or Codex or even one

of these models calculate what the ideal

context window you should define for

whatever model you're running. So, now

I'm just going to have it use Bright

Data CLI to tell me about the current

situation with Fable 5 and if it's going

to subscription plans. So, I'm going to

send it off and now we're going to watch

it use tools, specifically the Bright

Data CLI to go get it. But again,

because it's missing the Bright Data

skills, because there's no way to load

skills into Bionic, it doesn't use the

CLI as best possible. So, it's using the

CLI to understand what arguments it's

able to run. I also like, by the way,

how we could actually see the thinking.

The reasoning has been hidden or

truncated on Cloud and also on ChatGPT

on Codex. And here on Bionic, we're able

to actually see the thinking tokens.

Probably still summary, but still pretty

good. And there we go, July 18th,

emerging reports suggest that it's being

reintegrated into subscription plans

with a usage cap. So, overall, I think

this is really cool. I like LM Studio, I

like to run local models with it. Local

models are limited, they're smaller, it

really depends on your computer and even

if you have the most powerful computer

on the market, it's probably not going

to compare to anything you're using with

Fable or GPT 5.6 Soul. So, you have to

manage your expectations, but not only

manage those, manage how you expect to

work and get work done. It's all

possible. It's a workflow shift and I

suggest at least playing with this

understanding, getting to know open

source, getting to know open weights.

Now, what LM Studio Bionic is doing here

isn't that new. I've been actually

running open source models via Ollama

and Cloud Code for a few months now.

Bionic is essentially an agentic

harness. Cloud Code is an agentic

harness and Cloud Code is still my

favorite. It is very built out, it is

feature rich, but that being said, it is

not so efficient with small models

because it's built for large models. So,

the benefit here is if Bionic is built

with running smaller models, the smaller

context windows in mind, I believe it is

better positioned, obviously, for

running open weight models in the

future. That's my first take on LM

Studio Bionic. I've been running it for

last day since it came out on very small

coding and work tasks. It is capable. It

is missing some features, but I do like

LM Studio. I like what they've done and

I'm sure it will only get better. So, if

you're out of failure usage or out of

5.6 usage or you specifically run open

weights models, I suggest giving Bionic

a shot. For any questions or feedback,

drop in the comments below. If you

haven't done so already, subscribe to

the channel. It really helps me grow.

Thank you guys for watching and have a

great weekend.