Gemma 4 Has Landed!

Okay, so Google has just dropped Gemma 4

and this is four new models with

multimodality

thinking function calling the works and

honestly that alone would get me

covering this. But that's not even the

interesting part. The interesting part

is the license. Gemma 4 ships under an

Apache 2 license. Not a custom license

with weird restrictions with the whole

sort of open weights but don't compete

with us clauses. This is an actual real

Apache 2 license, which means for the

first time you can take Google's best

open model, modify it, fine-tune it,

deploy it commercially, do whatever you

want with it. No strings attached. And

when we combine that with inside these

models, we're talking about 128 experte

here, native audio, native vision,

built-in reasoning, all of that becomes

a pretty big deal. Okay, so let me give

you a quick orientation because there

are four models and the naming is a

little bit confusing here. Gemma 4 comes

in two tiers. You've got what they're

calling your workstation models. So this

is a 31 billion parameter dense model

and a 26 billion parameter mixture of

experts model with 4 billion parameters

active. And then you've got your edge

models. So this is the E2B and the E4B.

Now, these are tiny, efficient models

designed to run on phones, Raspberry

Pies, Jets and Nanos, and really pretty

much at the edge anywhere you need a

good quality model here. Now, I've

covered the Gemma line of models since

the original release. I covered Gemma 3

on the channel, and I know back then,

while a lot of people were very

impressed with it, they were kind of

frustrated with some of the things

around the license. So, you had this

capable model, but a license with enough

restrictions that a lot of people went

with Llama or went with Quen instead. So

the Apache 2.0 move here is Google

basically saying, "Okay, fine. We'll

play the same terms as some of the other

open model providers out there." And in

fact, as we're talking about this, some

of the other model open providers in

China are actually pulling back their

latest releases and not making them open

like they have in the past. So the other

big thing up front here is that Google

is saying that these are built from

Gemini 3 research. So basically the

architecture innovations that went into

some of their flagship commercial models

are slowly now trickling down into the

open weights models. So if you've been

running local models and I know a lot of

you have the landscape has kind of

settled into this pattern. We've kind of

gone past the llama models. We've now

got sort of quan mistral and they're all

sort of competing on benchmarks in this

sort of fixed parameter range for dense

models. But we've also seen, you know,

up until recently, most of these models

were text only or at best text plus

vision. If you want audio, you're kind

of bolting on whisper. You're bolting on

some external ASR pipeline. And often if

you wanted something like function

calling, you're kind of hoping that the

model cooperates with your prompt

template. So what Gemma 4 is doing here

is shipping all of that natively into a

single model family. vision, audio,

thinking, function calling, and all of

these four are actually built in from

the architecture level, not sort of

bolted on after the fact. All right, so

one of the key things that makes Gemma 4

better than the previous Gemma series is

that it now has the ability to do sort

of long chain of thought reasoning. And

we've seen clearly that this can improve

outputs and can get you better final

answers, etc. Now, not only can this

reason across text, but it can reason

across different modalities. So, it can

reason across images if you wanted to

basically pass in an image and make use

of that. And for the first time, you can

actually reason across audio. So, that

is also cool in here. Obviously, this

ability to do the long chain of thought

has improved a lot of the benchmarks

that are out there and they're getting

really strong results on the MMU Pro as

well as Sweetbench Pro. Along with the

reasoning comes function calling. So,

anything you want to do that's aantic,

you want to basically be using function

calling and tools. So, this has

integrated a lot of the research they

put into the function Gemma model which

they released at the end of last year.

But now this is both in the small models

and the bigger models. So a lot of

people will think that this is not that

new. But really the way people did this

in the past for doing this kind of

function calling was actually just

having the model to be better at

instruction following and then sort of

coaxing it into it. Gemma 4 actually has

the function calling baked into it from

scratch. So this is sort of optimized

for multi-turn agentic flows allows you

to do with multiple tools and it really

shows up in some of the agentic

benchmarks and tasks that you can do.

All right, I mentioned earlier in the

reasoning that the two smaller models,

unfortunately not all four models, but

the two smaller models actually have

audio support and that audio support is

a lot better than what we had in Gemma

3N and some of the previous Gemma models

that had audio support. This means that

you can do things like ASR and

transcription, but you can also do

speech to translated text support. So,

I'll show you that when we go through

the walkthrough. On top of this, the

audio encoder is not only better, but

it's just a lot smaller. So, this helps

a lot for anything that you want to do

at the edge with these models that

you're just not going to be using as

much device storage and memory. Another

thing of comparing Gemma 4 to say the

Gemma 3N series is to do with the image

encoder. The image encoder with those

Gemma 3N models. While it was good, it

really was a bit sort of old-fashioned

in the way that they did it. It didn't

handle things like aspect ratios well.

And because of that, you would often see

that it didn't do a great job for things

like OCR, etc. The Gemma 4 models

basically have native support for these

interled multi-image inputs. My guess is

from playing with it that it's probably

had a decent amount of sort of OCR and

document understanding training in

there. And because you can do that sort

of multi-image input, you can actually

do video here and have reasoning across

those multi-im images. So generally just

comparing the Gemma 4 against Gemma 3

and Gemma 3N, you've got a lot more

updates in both with the smaller models

supporting the audio and better

multimodality support. And whereas Gemma

3N only had a context window of 32K,

even for the small models on Gemma 4,

they've got a context window of 128K and

then 256K for the bigger models. All

right, so let's talk about some of these

architecture choices and the model sizes

themselves. So the mixture of experts

model is 26 billion total parameters,

but only 3.8 billion are active at any

time. Now they haven't gone for a huge

number of experts like we've seen some

of the other models go for recently.

They've got 128 of these sort of tiny

experts, eight being activated for each

token plus one sort of shared always on

expert. So if we compare that to the

Gemma 3 model which the largest model

was a 27 billion parameter dense model

obviously in that case you are using all

27 billion at the same time. So roughly

this is giving you sort of the

intelligence of a 27b model with the

compute costs of something around a 4B

model. Now this you can certainly run on

sort of consumer GPUs and I'm sure that

even as I'm recording this before it

comes out we will see this on Oama on

LLM studio etc. And Google themselves is

also releasing the QAT checkpoints

that's the quantized aware training