AI Engineer Europe 2026
Sovereign Escape Velocity: Ownership with Open Models — Gus Martins and Ian Ballantyne, Google DeepMind
Read the talk
Sovereign Escape Velocity: Running Gemma on Hardware You Control
Owning a model changes where data can go, which tasks are economical, and who operates the system. Gemma 4 illustrates those choices from phone skills to local translation agents.
From a talk by Gus Martins and Ian Ballantyne
Before you start: Familiarity with model inference, GPU memory, and API clients will help with the deployment discussion and Python example.
When proprietary data cannot leave
What happens when a model needs access to proprietary data that cannot leave your infrastructure? A hosted API may offer excellent capabilities, but it cannot satisfy that constraint by intelligence alone. You need control over where inference runs, access to the model itself, and the ability to customize it.
That is the division of responsibility between Gemini and Gemma 4 in Gus Martins and Ian Ballantyne’s Google DeepMind presentation. Martins, who works on Gemma, introduces the family released the previous Thursday, with further launch detail reserved for Omar’s keynote and Cassidy’s session. He presents Gemini as Google’s strongest, easiest starting point; Gemma supplies the control that a model accessed through Google’s hosted API cannot. The two families serve complementary deployment needs.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
What the effective sizes mean
The launch comprises four sizes. The smaller Gemma 4 E2B and Gemma 4 E4B target phones, IoT hardware, and other constrained devices. The E means effective: the name describes the effective parameter scale, rather than counting every parameter associated with the model.
Martins explains E2B as roughly a 5B-parameter model whose transformer computation has a much smaller footprint. Additional token-mapping parameters can reside outside accelerator memory. The model card describes these as per-layer embedding lookups. This is the useful distinction behind the naming: effective parameter counts are not gigabytes of memory. Actual memory requirements also depend on representation and runtime configuration; neither E2B nor E4B is a direct GPU-memory specification.
That arrangement is intended to make local execution practical on a Pixel or another phone. Both small models accept text, vision, and audio input, but produce text output. They also support thinking, coding, and function calling, so the phone deployment is not limited to a plain text chatbot.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Active computation, total weights, and useful capability
The larger options are Gemma 4 26B A4B, a mixture of experts, and Gemma 4 31B Dense. In the mixture-of-experts model, only part of the model participates in each token’s computation. Martins describes this through roughly 4B-sized working components. That figure refers to active computation, not the space needed to store the full model. The dense 31B is presented as the strongest model in the family.
Martins estimates that the two larger models stood fourth and seventh among open models on LM Arena at talk time, without assigning a rank to each. He also describes other models in the top twenty or thirty as at least two or three times larger, and sometimes twenty times larger. These are his leaderboard snapshot and size comparison, rather than a controlled task evaluation. The practical attraction is capability at a size that can fit more deployment environments; he cites coding, agentic work, and multilingual use as regular applications of the 31B model.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Enough capability for the task
Both larger models can run on a desktop or behind a server endpoint. Hosting them in a cloud environment still leaves room for ownership: the relevant question is who controls the model and its deployment, not whether the hardware sits under someone’s desk.
Choose for the work that needs doing. Martins explicitly concedes that these are not the most intelligent models available. Email summarization, coding assistance, document search, and interactions with documents do not necessarily require that distinction. A sufficiently capable smaller model can make those workloads accessible on less hardware.
Martins contrasts a 31B deployment on one GPU with unnamed competitors requiring 200 GB of memory and perhaps four or five GPUs. The comparison does not specify workload or configuration. The launch announcement supplies a narrower hardware condition: unquantized bfloat16 weights fit on an 80 GB H100. Fitting weights is a starting point for deployment sizing, not a measurement of serving throughput.
For an initial trial, Martins points to ai.dev, now redirecting to Google AI Studio. At the time of the talk, he describes free access to the 26B and 31B models alongside Gemini and Veo. He mentions combining vision, thinking, and code execution, although his planned demonstration of that combination is not shown.
The emphasis on LM Arena also has a specific purpose: its Elo scores reflect human preferences about responses. Academic benchmarks help characterize capabilities, while preference evaluation addresses how a model feels to query and how customers may experience its answers. Those are related but different questions.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Licensing is part of operational independence
Ownership also means being able to adapt a model and continue operating it without depending on a provider’s continued service or permission to access an endpoint. That is the operational meaning of sovereignty here: an institution can retain and run the model it has adopted.
Through Gemma 3, the family used a custom Gemma license that Martins describes as commercially friendly. But a custom license still requires interpretation. He illustrates the institutional friction with a hypothetical eighteen-month procurement review: even favorable terms can be difficult to adopt when legal teams must evaluate unfamiliar language. Gemma 4 moves to Apache 2.0, which Martins also announces as the direction for subsequent releases. That change does not retroactively change the licenses of earlier Gemma models.
The national examples span several generations. Martins mentions Gemma being used in parts of Ukraine’s services, a Bulgarian-language model based on Gemma 2, and work toward bringing that Bulgarian effort onto Gemma 4. He then describes a Brazilian adaptation of Gemma 3 for Portuguese. The Ukrainian implementation is not specified, and the language adaptations are not named in the talk.
Those adaptations raise a changing fine-tuning question. If the base model already performs strongly in a language, tooling may no longer be the main obstacle; finding a worthwhile improvement can be harder. Martins illustrates the tradeoff as substantial work for perhaps a one-percent gain, without specifying a metric. He reports top-two-or-three language rankings for the 31B model on LM Arena in many languages, but does not identify the languages. The decision is therefore whether a particular adaptation improves the institution’s tasks enough to justify the effort.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Four thresholds for offloading agent work
Ballantyne turns from model size to the amount of work agents generate. Agentic tasks consume tokens through repeated instructions, intermediate outputs, and subsequent actions. Taking ownership of inference can make that workload more controllable, especially when suitable hardware is already paid for. Citing OpenRouter’s December 2025 State of AI report, he identifies programming as one of the high-token categories when input and output are considered together. The report describes traffic through OpenRouter, rather than all AI usage.
A laptop GPU can potentially absorb document processing, data analysis, research, or suitable coding tasks. Ballantyne draws a boundary around that last category: he would not ask this model to redesign an entire application architecture. He favors specific instructions for refactoring, analysis, and generating small modular pieces of code. Those bounded jobs provide a more tractable unit of work to move onto personal hardware or a single GPU.
| Threshold | Question to answer |
|---|---|
| Capability | Can the model complete this specific task? |
| Hardware fit | Can the intended machine run it? |
| Latency | Can it respond within the task’s time budget? |
| Cost | Does owned or leased compute make economic sense? |
These thresholds have to hold together. Ballantyne describes user-facing work as potentially needing a response within a couple of seconds, whereas batch processing can tolerate a different schedule. Already-owned infrastructure, a planned hardware purchase, and leased GPU time each produce a different calculation. The useful question is which tasks you can fully offload and operate yourself.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
From phone input to a defined action
The mobile demonstration, credited to Cormac’s team, puts agent skills on a phone. A user can speak, provide images, show the surrounding world, or chat with the model. The model then consults the available skills: those can trigger device applications such as calendars and maps, or invoke skills defined by the developer. The mechanism connects multimodal input to a bounded set of actions the application exposes.
Ballantyne identifies reasoning about which action to take, followed by reliable calls to defined functions, as an improvement over the previous generation. The playground is Google AI Edge Gallery, available on iOS and Android. He tentatively identifies the demonstration as using the smaller 2B option; a 4B option is available for hardware with sufficient capacity.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Local inference changes the operating budget
Moving up to desktops and single GPUs brings the 26B and 31B models into consideration. But removing a per-token bill does not remove inference cost. GPU and NPU utilization consumes energy, so execution timing becomes a product decision. A response to a photograph may need to arrive immediately; other processing can run offline in the background while a phone charges overnight. Ownership makes scheduling and device utilization part of the application design.
For enterprise deployments, Ballantyne contrasts earlier choices involving models above 300B parameters and multiple GPUs with possible deployments on a single H100 or A100, and an L4 in some cases. Serving a team or company still requires a calculation for the actual workload: the accelerator name alone does not determine capacity or savings.
Customization adds another reason to operate the model yourself. Ballantyne introduces MedGemma as a medical model family that can work with private data under an organization’s control. He offers a hospital deployment on one or two GPUs as an illustrative possibility, not measured hospital-wide serving capacity. The enterprise opportunity combines specialization with control over where sensitive data is processed.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A local orchestrator fans out translation work
The desktop demonstration uses LM Studio to run the 26B model, the faster of the two larger options, with approximately 4B activated parameters. Ballantyne reports approximately 26 GB of RAM usage, including context, on an M4 Mac with approximately 48 GB of unified memory. This is a concrete reminder that active parameter count and total runtime memory are different quantities.
The workflow starts in a terminal:
- An orchestrator receives the Gemma 4 announcement as the source document.
- It distributes translation work to sub-agents, each represented by a separate window.
- The sub-agents generate translations into different languages on the same local machine.
The initial wait includes processing in LM Studio before translations begin appearing. Multiple agent windows here represent separate pieces of work, not separate physical machines.
The intended final step is to collect the translations and generate a small webpage containing the results. Ballantyne describes that aggregation prospectively; the demonstration establishes translation generation but does not clearly establish the final webpage’s contents. He offers file processing and additional analysis as other tasks that could use the same local orchestration pattern.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Change the endpoint, then evaluate the workflow
The first integration step is small: point an existing OpenAI-compatible client at LM Studio or Ollama, then select the locally served Gemma model. Current compatibility documentation supports this endpoint-and-model pattern for supported APIs; it does not imply complete feature equivalence. In Python, the translation task can start with a client like this, using the model identifier exposed by the local server:
python
import os
from pathlib import Path
from openai import OpenAI
client = OpenAI(
base_url=os.environ["LOCAL_OPENAI_BASE_URL"],
api_key=os.environ.get("LOCAL_OPENAI_API_KEY", "local"),
)
announcement = Path("announcement.txt").read_text(encoding="utf-8")
response = client.chat.completions.create(
model=os.environ["LOCAL_GEMMA_MODEL"],
messages=[
{
"role": "user",
"content": (
"Translate this Gemma 4 announcement into Portuguese. "
"Preserve model names and technical identifiers. "
"Return only the translation.\n\n" + announcement
),
}
],
)
translation = response.choices[0].message.content
if not translation:
raise RuntimeError("The model returned no translation text.")
Path("announcement.pt.txt").write_text(translation, encoding="utf-8")
This makes one translation request through the local endpoint. An orchestrator can assign the same source document to separate language tasks, as in the demonstration.
The next steps are empirical:
- Drop the model into an existing workflow. Observe which tasks it handles well.
- Separate tuning needs from capability limits. Some failures may be addressable through adaptation; others may reflect task complexity beyond the model.
- Strengthen the evaluation suite for your tasks. General benchmarks help identify candidates, but success depends on the work your application actually asks the model to perform.
The endpoint change gets the experiment started. Task evaluation determines whether the substitution is useful.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The responsibilities that come with control
Self-hosting transfers uptime, downtime, maintenance, and ongoing operating costs to the owner. Buying hardware also adds upfront capital expenditure. Those obligations belong in the same decision as model quality and inference cost; they are part of what it means to control the service.
Mobile deployment adds a support matrix: which devices are supported, which accelerators they contain, and how much RAM is available. In exchange, the application can work offline or process private information that never leaves the user’s device. Enterprise scaling requires the same attention to infrastructure sizing and continuing operating costs at a larger scope.
Ballantyne closes by encouraging experimentation with concrete tasks, using benchmarks as signals of what may be feasible. The feedback he asks for is equally concrete: how deployments behave, what fine-tuning changes, and which problems emerge. Those observations are what turn access to an open model into a system an organization can confidently operate.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
The original announcement covering the four launch sizes, Apache 2.0 licensing, capabilities and hardware options.
Installable playground for experimenting with local mobile models, multimodal inputs and agent skills.
December 2025 study of more than 100 trillion tokens, including programming workloads and input/output sequence lengths.
Configure an existing client to use a locally served model through supported OpenAI-compatible endpoints.
Current documentation for connecting OpenAI-compatible applications to Ollama.
Google's developer overview of the MedGemma medical model family.
Further reading
Case study of adapting Gemma 3 4B to Brazilian Portuguese through continuous pretraining.
INSAIT's Bulgarian-language Gemma 2 adaptation, with model details and usage instructions.
Updates since the talk
- Gemma 4 model cardDocumentation
Current architecture tables, parameter definitions and evaluations, including the expanded five-model family.
Read the complete timestamped transcript
- 0:00
[upbeat music] Hi, everyone.
- 0:16
Uh, can you hear me? Yes, you can hear me. Hi. Sorry, sorry, one minute late. I'll try to do my best to finish earlier so my friend can do, uh, pretty cool demos for you.
- 0:25
I'm Gus. This is Ian. We are from the, uh, Google DeepMind, and I'm specifically work on the Gemma product. Do any of you know what Gemma models are? Okay, perfect.
- 0:37
Perfect. Thank you very much. Uh, so today we're going to talk a little bit about ownership and open models. Uh, and, uh, well, you know who we are. But the idea is, uh, last Thursday we released our new, uh, family of models, Gemma 4, and I'm going to talk a little bit about them.
- 0:53
Uh, there is going to be more information tomorrow in the keynote by Omar, and there's another talk by Cassidy also tomorrow that you'll go into even more details. We are going to tell a little bit of the story, but the story is a little bit bigger.
- 1:06
We'll try our best here. Uh, so why does it matter? Uh, if you ask me, I work for Google, of course if you ask me what's the best model for you to try, the easiest one, I will answer for you, Gemini.
- 1:18
Gemini is the best model we have, pretty strong, multimodal, can do all kinds of things. But then there is, uh, there's more to this story than just having the strongest model possible.
- 1:31
In some situations, you want to own the model. You want to be able to run on your own hardware. You want to customize it. You want to be able to send your proprietary data that cannot leave your infrastructure.
- 1:45
So there are many situations where even the best proprietary model will not be able to help you directly. That's when you might need an open model. That's where Gemma comes in.
- 1:56
So when you think, why does Google have two family of models? Because they complement each other. So Gemini is the m-most intelligent one, can do a lot of cool stuff, but it's hosted in Google servers.
- 2:07
You need the API to access. If you need more control and access, you need an open model. That's why we have Gemma. That's why, uh, and, and we are very proud that the quality is very, very strong.
- 2:18
We're going to go some details later. But the idea is you would be able to do a lot of cool stuff with it. Among the launches we did, we released four sizes.
- 2:26
Uh, two are target to mobile, let's say like that, or, uh, or IoT or smaller devices. It's a E2B and a E4B. Uh, these names is a little bit weird.
- 2:37
We are the only ones that use this name. Uh, and the E stands for effective. And the idea here is, uh, the model, uh, uses as much as 2B, uh, what a 2B model would use as memory, but it's larger than that.
- 2:53
The 2B is around 5B, uh, parameters. But then you say, "Oh, but where is this other 3B in memory?" The f- the fun fact is that they are not really parameters from the transformers.
- 3:05
They are, uh, like mapping, uh, tokens. So you can leave them in other memory. So what you really need on your GPU memory is the two billion or the four billion.
- 3:16
Why do we do that? So that you can run these models on a phone, on a Pixel phone or b- any phone you have there. You can run these models, and they're very strong.
- 3:24
The E2B and E4B, both of them have, uh, text, uh, vision, and audio input, and they do only text output. They can do thinking. They can do coding, function calling, all these kind of cool things.
- 3:37
These all run on your phone right now. You could download it right now, right?
- 3:42
We also have other two models, which are the larger ones. We have a 26B and a [REDACTED:generic_id]. The 26 is a mixture of experts, which means that, uh, it's as if we had many other models working together where, where each of the...
- 3:56
where each one of those are like a 4B model. Why does it matter? Because it, it has twenty-six billion parameters, but it needs a space of a four billion param to do the work.
- 4:07
And this makes, uh, makes it accessible to way more hardware, to way more people, and it's still pretty strong. But our strongest model is the [REDACTED:generic_id] Dense, which is thirty-one billion parameters model, and this is really, really strong.
- 4:20
If we look into our, uh, Elo score on LM Arena, you can see that both our models are-- They are, I guess now they are fourth and seventh as the leads on open source models, open models.
- 4:34
And if you compare them to maybe the top twenty, thirty, all of them are at least twice, three times larger than our models. In some cases, twenty times larger.
- 4:45
So we are talking about a disproportionate amount of intelligence per size. So our [REDACTED:generic_id] model is the one I use very regularly. It can do basically anything from coding, agentic everything, multilingual, all of that.
- 4:59
So I strongly recommend you try those. They are so strong that they are, uh, o-both of them are really good to use on your, on a, as a cloud deployed model.
- 5:08
They can run on your desktop, but if you use on your server as your endpoint to do s- uh, your work, they are pretty good. And you ask, "Oh, is this the most intelligent model?"
- 5:18
No, it isn't. I'm going-- I'm, I'm very biased, and I love them, but I know the capabilities. But the question is, do you need the most intelligent model of the planet to summarize your email, to do some more menial tasks, to help you code, to do some agentic capabilities that are s- uh, searching and interacting with docs?
- 5:37
Probably not. That's why these models are so strong, because they're cheaper. They're very strong, but they're cheaper to run. They require way less hardware. A [REDACTED:generic_id] running one GPU.
- 5:48
The competitors need two hundred gigabytes of memory, which would be maybe four or five GPUs. So you can see that the price here is really, really different.
- 5:59
One easy place for you to try these models is on ai.dev, where you can try Gemini models, Veo, all the other models. But Gem- Gemma are there, both 26 and [REDACTED:generic_id].
- 6:08
You can try right now. They are free. You can play with it. Uh, and they can do-- I was, I was going to show a demo, but I, I can't now.
- 6:14
But they can do some cool stuff, which is vision plus thinking plus code execution all at the same time, right? Uh, I'll try to post something about this later.
- 6:24
But the idea is you can play with the models pretty easily and right there, r- not now. Finish, let's finish the, the, the talk and then you play with it.
- 6:32
Uh, and as I was saying, the in-the intelligence per parameter that these models bring is pretty good. It's very, very strong. And if we, we use the Elo score for LLM Arena because it's a benchmark that's a, a person's preference, right?
- 6:45
We can look into academic benchmarks. They are very, very strong. But the, how the model responds to your queries, that's very important, right? That's how your customers will see, how you will see and, and interact with it.
- 6:59
So this is why this is so important. And
- 7:03
why does all this matter? One of the reasons that we care so much is because, uh, you want to, you want to-- the user to have ownership. And more than that, we are enabling sovereignty.
- 7:16
And sovereignty me-means in terms of you own the model, and you can adapt your use cases, and you, you are not susceptible to, I don't know, loss of service or for some kind of, uh, someone saying, "No, no, you cannot use this model anymore."
- 7:30
It's all available to you. And one of the changes we made last year until Gemma 3 and others, we had our specific license, a Gemma license, which is pretty good, commercial-friendly and all.
- 7:42
But there's a problem. If you have a custom license, I don't know if you have any lawyers here. Uh, if I tell you, "Oh, we have this custom license," your lawyers will look at me with that face that, "I hate you, Gus."
- 7:53
And then they will spend like 18 month doing procurement process to understand the license and trying to change, and that never works. So it's pretty hard for sovereignty institutions to adopt this kind of thing.
- 8:04
That's why we moved to a pass 2.0 for, for Gemma 4 and going forward. That makes- [audience member sneezes] Thank you. And that makes our life mu- uh, your life much easier to, to convince your legal department, let's say like that, that, "Look, we own this model.
- 8:18
We can use." So this is pretty important, and it enables many, many, uh, sovereignty institution to use our models. We have some examples. For example, ki- for example, uh, Ukraine used Gemma to, in parts of their inf-- uh, uh, services.
- 8:31
We have a, a one version of the Gemma model that was fine-tuned for Bulgarian. It was their, uh, LLM for the country. That was based on Gemma 2. We are working to make sure they use Gemma 4 now.
- 8:42
We also have a Brazilian version that is based on Gemma 3. It was fine-tuned for Portuguese. And the challenge of these models today is that they-- if you want to fine-tune Gemma to a specific language, it's becoming very hard to do that.
- 8:56
And the problem is hard because not the f-the tooling or anything. It's because the model is pretty strong on those languages already. So any gains you try to have, you might not get there.
- 9:06
So you might spend a lot of time to get like one percent. And then maybe, I don't know if it's the best use of your, uh, time. So this is good and bad at the same time 'cause-- but it's good that you can automatically use in many language.
- 9:18
You can try right now. And if you're going to, uh, the LLM Arena for languages, in many language, they are top two, three. And look, it's a [REDACTED:generic_id] model.
- 9:28
It's very, very small, right? So this is pretty good.
- 9:32
That being said, uh, I will let my colleague continue and show some demos. [microphone feedback]
- 9:38
Thank you, Gus. So, uh, one thing that I think is really important about these models is that when you think about using open models, you think about like using proprietary models.
- 9:48
We're move-- We're seeing a shift now to more kind of agentic capabilities and the kind of tasks that we're trying to do. And with that comes a cost in tokens and token generation.
- 9:57
So one benefit of, uh, taking ownership of the models is your ability to control, or in cases where you have sunken hardware cost, uh, to be able to iterate on top of that.
- 10:07
Uh, this graph on the right-hand side is from the State of AI report that OpenRoute did, and it shows the, uh... It's a bit small for you on this diagram, but have a look at that link.
- 10:16
Uh, it shows the different types of tasks that people are doing through OpenRoute at the moment. And you'll see the, the, the one that's about here, this one here is programming is right in the middle, and this is kind-- among some of the highest tasks in terms of token generation, both input and output combined.
- 10:33
So the more we have agents work and do these kind of tasks for us that have very high token generation costs, that's when you start to get more benefit from being able to take control of that i-in itself.
- 10:45
So if, for instance, you have a laptop that is capable of doing a particular task that you need to be doing, like processing a document or analyzing some data or doing some research or, in the cases, uh, Gus talked about, doing some coding that's suitable for that, then you have a GPU that you can take advantage to
- 11:02
do some of that stuff. Now, similarly, similar to what Gus said about, you know, we don't necessarily-- We still have frontier models for doing the best possible things. I wouldn't get this model to do like a, you know, a full systems architecture and redesign of your application, right?
- 11:16
It's not kind of for that. But what it is very good at doing is following very specific instructions about doing things like refactoring, analyzing, uh, generating code in, uh, in small modular bits, and you can offload a chunk of work in that style to these kind of models to be able to do that, whether it's on a
- 11:33
single GPU or on your own personal hardware. Um, and the way that we kind of think about this is like a, a set of thresholds. Like, when do we get to the point where these models are capable of doing the task, but then they also fit on the right hardware, that they also, uh, can do it with
- 11:51
the right amount of latency, depending on the, the-- if it's a task for a user, it needs to happen in a couple seconds. If it's a task where you're doing things like batch processing, you maybe have slightly different thresholds for like what needs to be done.
- 12:04
Um, and then also what the cost of actually doing that is. So if you have a sunken cost in terms of like infrastructure that you already own or that you're prepared to, uh, outlay and then operating on that, or whether you're leasing like GPU time or something else like that.
- 12:18
So these are gonna be very specific to the task that you're trying to achieve. But what, uh, what you can do with open models is you can think very carefully about like what-- which of these tasks can I f- can I fully offload or can I fully own compared to relying just on using the best possible models
- 12:33
to do, uh, that in the cloud. And, uh, an example, so Gus talked about the different types of hardware that can run these things now. I'm just gonna run this little demo in the side at the moment.
- 12:45
So we now have models that will work directly on mobile and edge devices. Uh, this example here, um, was built by Cormac's team, um, is a set of agent skills that the model is running on a phone.
- 12:58
So, uh, I'm gonna mute the microphone for that. Um, so you can talk to the model, you can show it images, uh, you can show it the world around you, and you can prompt it and chat to it.
- 13:09
And what this one is showing is that it can look through a set of skills that it has about things on the phone. So either it can take actions on the device itself, so trigger other applications, like trigger calendar apps, trigger maps apps, or you can kind of define your own skill sets.
- 13:25
And what's different now with the Gemma 4 models than we saw for the previous generation is that it's able to reason about what actions it needs to take and reliably, uh, make those function calls defined.
- 13:36
So what this app will allow you to do is, is kind of acts as like a playground. Uh, so this is Google AI Edge Gallery, and you can find it on iOS and Android.
- 13:44
And you can experiment to see what the models of this size are actually able to do. So I think this is the two billion parameter model, but there's also the four billion parameter model, depending on, uh, the size of your hardware.
- 13:56
And when we get to desktops and single GPUs, as Gus mentioned, that's where you can use the, uh, the 26 and the [REDACTED:generic_id] models, again, on your local hardware, and I'll show you how to do that in a minute.
- 14:07
But there's one kind of key point here is, well, whereas we're not paying for, uh, the price of these agents or models within tokens, we're actually paying for them in terms of energy cost if we think about it.
- 14:18
Because now you're thinking about utilization of GPUs, you're thinking about utilization of MPUs on the hardware itself. Uh, when are you gonna do these tasks? Does the user need to get a response right now when you're taking a picture of something?
- 14:30
Or is it something that you can process offline as a background task when they plug their phone in at night? So the-- what I'm trying to say here is that the thresholds and how you think about the usage of these models kind of shifts when you come to on-device or ownership because you think more about how they're
- 14:46
being executed and why they're being executed. Um, yeah, perfect. Uh, and similarly, on the enterprise side, if you don't have a piece of hardware that can, you know, run the thirty-one billion parameter model, you can now be thinking about scaling that down.
- 15:01
So maybe if you wanted to use a three hundred plus, uh, billion parameter model before, you might have need multiple GPUs. Now you can think about using a single, uh, H100 or A100 or even in some cases like an L4.
- 15:15
And then the costs obviously related to that also kind of go down. So again, it's a calculation that you'll have to do depending on your use cases, but there's ways that you could scale, for instance, running one of these models to serve, you know, a small team or to serve a company, um, depending on, uh, what you're
- 15:30
trying to do. And the f-- and the f-- and the final point is that you also have the fine-tuning component too, which is that because these models can be customized, you can deploy your own version of it.
- 15:40
So for instance, we have a variant of Gemma models called MedGemma, which is specialized for medical use cases. So if you wanted to have something that would operate on private data that you can control yourself, you can now d-- re-- feasibly deploy this to, like, one or maybe two GPUs, uh, to run that for, I don't know,
- 15:57
like a whole hospital, for instance. So these are kind of worth considering for the enterprise case. Uh, I'm gonna jump straight to demos now. Um, I've shown you some demos on the phone.
- 16:08
I'm gonna show you a quick demo, uh, here. Who-- Quick show of hands, who's ever used, uh, a, a tool called LM Studio?
- 16:17
Okay. Uh, just under half people. So LM Studio is a way that you can play around with local models. And I have here-- I have-- This is the 26B model, so this is our faster of the, the two larger models with four billion activated parameters.
- 16:32
And I'm-- At the moment, including the context, it's probably about twenty-six gigabytes in RAM. And this is an M4 Mac. So I've got unified memory. I've got up to about forty-eight gigabytes.
- 16:42
Uh, so I can run it on this machine. And I'm just gonna run this terminal right here. Let's give that a go. Oops. Pre-showing my demo. Let's try that again.
- 16:54
There we go. So I'm just gonna run a little process where I'm going to do some quick-- a trick-- quick translation on my device. So what it's gonna do is I've got an orchestrator on this side here, uh, which is going to hopefully kick off my agent in a minute.
- 17:12
Let's make sure we are loaded. Let's see what LM Studio is doing. Yeah, it's just processing at the moment. And then it's gonna farm out, uh, this translation to all of these different, uh, uh, windows, and each one of them represents a different sub-agent.
- 17:26
So this is running on my device, uh, and it's going to basically execute all these translations in one go. So I've given it, like, the Gemma 4 announcement, and I just wanna s-- wanna translate to all these different languages.
- 17:38
So you'll see in a second it should hopefully send it over there.
- 17:43
Three, two, one. And hopefully we should be generating translations in a second. There we go. So, so you can imagine doing any kind of agentic task on your local machine.
- 17:54
Uh, you could have it, like, processing files. You could have it doing, uh, additional analysis. And hopefully, what you'll see in a minute is it will be able to compile all these back, and then it will generate me a quick webpage, and then you can see the results of your translation.
- 18:08
There you go. So there's the multilinguality of the model there as well.
- 18:14
Thank you. Uh, [clapping] Right. Uh, so in the interest of time, uh, I just want to say that the, the main next step for exploring and trying out these models is as simple as this code on the right-hand side.
- 18:32
Um, you can take any OpenAI-compatible interface that you've got, and you can point it at a service like Ollama or LM, uh, LM, uh, LM Studio, and you can just pick out the Gemma model, and that's all you need to change code-wise to at least, uh, try it out.
- 18:47
So the first thing we recommend you do is to drop it into existing workflows that you have to then see what the model can handle. Like, what is it working well at?
- 18:55
What would it need tuning for? What is kind of out of its depth in terms of, like, the complexity of the task? Um, next is to kind of bolster your evaluation suites because, you know, benchmarks are great and everything for just saying what general capabilities are.
- 19:08
But the reality is that how good the model is depends on how well it does on your task and not anybody else's task. Um, the other thing I mentioned very briefly is thinking about how you actually serve these models in the end.
- 19:20
So if you need to run your own GPU, you need to host it. Yes, you're in control of, like, uptan-- uptime and downtime, but then there's, like, maintenance costs and there's stuff like that.
- 19:29
So you have to be... You have to consider that as, like, one of the factors, like the ongoing costs as well, as well as any upfront CapEx costs if you buy infrastructure or hardware to do that too.
- 19:39
Um, on mobile devices, for instance, you have to think about, like, if I'm gonna offload stuff to a phone, like, what am I supporting? What accelerators do they have?
- 19:47
What size RAM do they have? So the conversation becomes a little bit more complex, but then there's a whole heap of things that you can unlock, like working offline or working on users' private data that never leaves their device.
- 19:58
Um, and finally, if you wanna scale this up to enterprise levels, you have to think then again about, like, the kind of infrastructure that you're running on and what the ongoing costs are of that as well.
- 20:08
But it does kind of unlock that. So, uh, with that, the summary is that you can use these models in pretty much any way you can think about. Experiment what kind of tasks are possible with it.
- 20:19
Use some of the benchmarks to kind of give you an indication of, like, what's feasible. But, uh, really, we want to hear your feedback and how you get on with these and, and how you fine-tune them and what kind of things you run into, and, uh, we wanna help you on that journey as well.
- 20:33
So with that, thank you very much. [audience applauding] [upbeat music]