Ian Ballantyne and Gus Martins (Google DeepMind) discuss Gemma 4's 31B open model, which ranks fourth on the LM Arena open model leaderboard despite being much smaller than neighboring models and running on a single GPU. Topics include size efficiency (E2B/E4B models using only 2B/4B of GPU memory despite more parameters, enabling on-device runs like a Pixel phone), a medical variant deployable on two GPUs for an entire hospital, parallel multi-agent workloads on an M4 Mac via LM Studio, and the license shift from the custom Gemma license to Apache 2.0, which eases adoption by sovereign institutions such as those in Ukraine, Bulgaria, and Brazil. # Channel Identity Observations: Video `01a15ca9-0c60-7d88-849f-439e338dff85` ## Video Identity - **Title:** *Sovereign Escape Velocity: Ownership w Open Models — Gus Martins, & Ian Ballantyne, Google DeepMind* - **Format:** Conference/event talk with live demos, presented by two Google DeepMind employees. ## Explicit Self-Description Evidence - **Speakers identify themselves and affiliation directly:** > "I'm Gus. This is Ian. We are from the uh Google deep mind and I'm specifically I work on the Gemma product." - **Claimed role:** Gus positions himself as a product person on Gemma; Ian presents demos. The talk is framed as part of a larger event program: > "there's going to be more information tomorrow in the keynote by Omar and there's another talk by CassD also tomorrow" This indicates the channel hosts event talks rather than producing standalone content. ## Demonstrated Scope (Observable Presentation) - **Subject matter:** Open-weight AI models — specifically the Gemma model family (Gemma 4 launch), positioned relative to Google's proprietary Gemini models: > "Gemini is the most intelligent one… If you need more control and access, you need an open model. That's why we have Gemma." - **Recurring themes demonstrated in the talk:** - **Model ownership / sovereignty** — running models on own hardware, customizing, keeping data in-infrastructure, and licensing (move to Apache 2.0): > "we move to a pass 2.0… for Gemma 4 and going forward" - **Technical depth** — parameter counts (E2B/E4B, 26B MoE, 31B dense), memory requirements, Elo scores, multilingual fine-tuning examples (Ukraine, Bulgaria, Brazil). - **Practical deployment guidance** — hardware thresholds, cost tradeoffs, evaluation advice, serving considerations. - **Live demonstrations** — on-device mobile agent skills via "Google AI gallery" app, and a local multi-agent translation demo using LM Studio on an M4 Mac. ## Intended Audience (Observable) - **Developers and technical practitioners:** frequent hands-on calls to action — "You could download it right now", "try these models… on AI.dev… They are free", "point it at a service like Olama or LM Studio". - **Enterprise / institutional adopters:** discussion of procurement, legal departments, fine-tuning, deployment on H100/A100/L4 GPUs, and sovereign/national use cases (Bulgaria's national LLM, Brazil's Portuguese variant, MedGemma for hospitals). - **Audience interaction confirms a live technical crowd:** > "Do any of you know what Gemma models are? Okay, perfect." > "Who's ever used uh a tool called LM Studio? Okay, just under half people." ## Tone and Style (Demonstrated) - **Informal, candid presentation:** speaker openly acknowledges bias and limits — "I'm I'm very biased and I love them but I know the capabilities"; admits a demo couldn't run ("I was going to show a demo but I can't now"). - **Balanced framing:** repeatedly distinguishes open models from frontier models rather than claiming universal superiority: > "is this the most intelligent model? No it isn't." > "I wouldn't get this model to do like a… full systems architecture and redesign of your application." - **Feedback-seeking posture:** the closing invites community engagement: > "really we want to hear your feedback and how you get on with these and how you fine-tune them." ## Notes for Channel-Level Synthesis - This video demonstrates a channel featuring **industry practitioner talks from major AI labs** (here, Google DeepMind), with vendor-identified speakers presenting product content — claims about model quality come from the vendor itself. - No curated metadata was supplied; all observations derive from title and transcript only. - No private motivations were inferred; all statements above are directly quoted or precisely described from the transcript. # Delivery Observation: "Sovereign Escape Velocity" — Gus Martins & Ian Ballantyne ## Overall Format - **Two-speaker talk** with a clear handoff: Gus covers strategy/product context (~first 60% of transcript), Ian covers use cases and live demos (~final 40%). - Gus opens with logistics and an apology: *"Sorry we're 1 minute late. I'll try to do my best to finish earlier so my friend can do a pretty cool demos for you."* - The talk closes with Ian's call for feedback: *"really we want to hear your feedback and how you get on with these... So, with that, thank you very much."* --- ## Speaker 1: Gus Martins ### Tone & Style - **Conversational, self-deprecating, and candid.** Frequent fillers ("uh," "you know," "like that") and informal asides. Example of transparency about bias: *"I'm I'm very biased and I love them but I know the capabilities."* - **Direct rhetorical questions to the audience:** *"Do any of you know what Gemma models are? Okay, perfect."* - Uses humor to illustrate a point about licensing: *"your lawyers will look at me with that face that I hate you, Gus, and then they will spend like 18 months doing procurement process."* ### Vocabulary & Recurring Phrases - **"Pretty" as an intensifier** — very frequent: "pretty good," "pretty strong," "pretty easily," "pretty important." At least 8 instances. - **"Very" doubled for emphasis:** *"the quality is very very strong"*, *"It's very very small"*, *"very very strong."* - **"Cool stuff / cool things"** as a repeated catch-all descriptor: *"do all kinds of cool things"*, *"you would be able to do a lot of cool stuff with it."* - **"Strong" as the default praise word** — used roughly 8+ times ("very strong," "still pretty strong," "really really strong," "they're very strong, but they're cheaper to run"). ### Teaching Habits - **Self-question-and-answer structure:** poses a question, then answers it. - *"And you ask is this the most intelligent model? No it isn't."* - *"Why does Google have two family of models? Because they complement each other."* - *"But then you say oh but where is these other 3B in memory"* - **Concrete memory/parameter math** to explain architecture: *"The 2B is around 5B... parameters. But... they are not really parameters from the transformers. They are uh like mapping uh tokens."* - **Explicit honesty about limitations**, framing it as a use-case argument: *"do you need the most intelligent model of the planet to summarize your mail, to do some more minial tasks... Probably not."* - **Audience-directed action prompts:** *"You could download it right now."*, *"I strongly recommend you try those."* ### Self-Corrections Mid-Sentence - Noticeable mid-utterance repairs: *"G gem are there"*, *"not now finish let's finish the the talk and then you play with it"*, *"the problem is hard because not the the tooling or anything because the model is pretty strong."* --- ## Speaker 2: Ian Ballantyne ### Tone & Style - **More structured and hedged** than Gus, with frequent softening qualifiers: "kind of," "kind of," "like," "a bit." - *"it's a bit small for you on this diagram but have a look at that link"* - *"it's not kind of for that but what it is very good at doing is..."* - Repeated framing phrase **"kind of"** — very high frequency throughout his segment. ### Vocabulary & Recurring Phrases - **"Kind of" as a hedge** — appears 20+ times (e.g., "kind of agentic capabilities," "kind of a full systems architecture," "kind of worth considering"). - **Framework vocabulary:** "thresholds," "offload," "utilization," "capex costs," "use cases." - **"So" as a discourse connector** starting many sentences, especially during demos: *"So, I'm just going to run a little process..."*, *"So you can imagine doing any kind of aentic task..."* ### Teaching Habits - **Thresholds/framework framing:** *"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."* - **Practical migration advice** given as a minimal-change instruction: *"you can take any open AI compatible interface that you've got and you can point it at a service like Olama or LM LM LM Studio... that's all you need to change codewise."* - **Audience interaction for calibration:** *"Who quick show of hands. Who's ever used uh a tool called LM Studio? Okay, just under half people."* - **Honest framing of costs/benefits:** acknowledges *"maintenance costs," "ongoing costs," "upfront capex costs"* of self-hosting. ### Live Demo Delivery - Narrates troubleshooting in real time: *"Oops. Pre-showing my demo. Let's try that again. There we go."* - Countdown-style delivery: *"3 two one and hopefully we should be generating translations a second. There we go."* --- ## Shared / Cross-Speaker Patterns - **Both qualify model capability honestly** rather than overclaiming: - Gus: *"is this the most intelligent model? No it isn't."* - Ian: *"I wouldn't get this model to do like a you know a full systems architecture and redesign of your application."* - **Both use concrete hardware/cost comparisons:** Gus cites *"a 31B running one GPU. The competitors need 200 GB of memory which would be maybe four or five GPUs"*; Ian cites *"26 GB in RAM. And this is an M4 Mac."* - **Both point to national/sovereign examples:** Ukraine, Bulgarian and Portuguese fine-tunes (Gus); MedGemma for hospitals (Ian). ## Uncertainty Notes - Speaker attribution relies on the handoff marker *"I will let my colleague continue and show some demos"* and style shifts; segment boundaries are inferred, not labeled with names per line. - Timestamps are not verifiable beyond bracketed transcript markers; no external claims (e.g., "fourth and seventh" ELO ranking) are confirmed — they are reported as speaker claims only. # Observation Record: Sovereign Escape Velocity — Ownership with Open Models **Speakers:** Gus Martins and Ian Ballantyne (Google DeepMind; Gus works on the Gemma product). Speaker identification is explicit in the transcript ("I'm Gus. This is Ian"). --- ## Gus Martins — Gemma 4 Overview & Open-Model Rationale ### Why open models exist alongside Gemini - Claims **Gemini is Google's best model** ("If you ask me what's the best model… I will answer Gemini"), but notes it's **hosted on Google servers and requires an API**. - Argues situations exist where "even the best proprietary model will not be able to help you directly": needing to **run on your own hardware**, **customize**, or keep **proprietary data inside your infrastructure**. - Framing: Gemini and Gemma are **complementary families** — Gemini for maximum intelligence, Gemma for **control and access**. ### Gemma 4 release details - Four sizes released "last Thursday": - **E2B / E4B** — targeted at mobile/IoT/smaller devices; "E" stands for **effective**. The E2B uses memory like a 2B model but is ~5B parameters; the extra parameters are token mappings (not transformer parameters) that can live in other memory, so only ~2B/4B occupies GPU memory. Both have **text, vision, and audio input, text-only output**, plus thinking, coding, function calling — runnable "on your phone right now." - **26B (mixture of experts)** — "as if we had many other models working together," each expert ~4B; needs the working space of a 4B model. - **31B dense** — described as the strongest; the model he uses regularly; capable of coding, agentic tasks, multilingual use. ### Performance and cost claims - On **LM Arena ELO scores**, the models are (at time of talk) roughly **4th and 7th among open models**, while competing top-20/30 models are "at least two-three times larger… in some cases 20 times larger" — a claim of **disproportionate intelligence per size**. - Honest qualification: **"Is this the most intelligent model? No it isn't."** He states his bias ("I'm very biased and I love them") but argues most tasks (summarizing mail, agentic search, coding help) don't need the most intelligent model. - Cost comparison: 31B runs on **one GPU**; competitors need "maybe four or five GPUs" (200 GB memory). - Usable free on **AI.dev**; capable of vision + thinking + code execution simultaneously. ### Sovereignty and licensing - Ownership framing: users are **not susceptible to loss of service** or a vendor revoking access. - **License change:** Gemma previously used a custom Gemma license (commercially friendly, but he argues custom licenses trigger 18-month legal/procurement pain, especially for sovereign institutions); **Gemma 4 moves to Apache 2.0**. - Sovereign deployments cited: **Ukraine** using Gemma in services; **Bulgaria's** country LLM fine-tuned from Gemma 2; a **Brazilian Portuguese** version based on Gemma 3. - Tension acknowledged: fine-tuning for specific languages is now **hard** — not because of tooling, but because the model is already strong in many languages, so gains may be ~1% and possibly not worth the time. "This is good and bad at the same time." - On multilingual benchmarks, the 31B is "top two or three" in many languages on LM Arena. --- ## Ian Ballantyne — Ownership Economics, Thresholds, and Demos ### Token-cost rationale for ownership - Shift toward **agentic capabilities** raises token generation costs. Cites the **OpenRouter State of AI report** graph showing **programming** among the highest token-generation tasks (input + output combined). - Ownership lets you exploit **sunken hardware costs** (e.g., your own laptop/GPU) instead of paying per token. ### What the models are (and aren't) good for - Explicit limitation: **"I wouldn't get this model to do a full systems architecture and redesign of your application."** - Good at: **following specific instructions** — refactoring, analyzing, generating small modular code — offloadable to a single GPU or personal hardware. - Proposes a **thresholds framework**: capability + hardware fit + latency requirement (seconds for user-facing vs. relaxed for batch) + cost model (owned infrastructure vs. leased GPU time), all **task-specific**. ### On-device cost model shift - Key point: with local models you pay in **energy costs**, not tokens — GPU/NPU utilization, and **when** tasks run (e.g., defer to overnight charging rather than real-time). ### Enterprise scaling - Where 300B+ parameter models previously needed multiple GPUs, Gemma 4's larger sizes can run on **a single H100, A100, or sometimes L4**, reducing costs. - **Fine-tuning/customization** enables private deployments; cites **MedGemma** as a medical-specialized variant, e.g., serving "a whole hospital" on one or two GPUs with private data. ### Demos - **Mobile agent demo** (built by "Coremax team," in Google AI Gallery on iOS/Android): on-device agent skills — triggering calendar/maps apps, custom skills, multimodal input. Claims Gemma 4 improves on previous generation in **reasoning about which actions to take and reliably making function calls**. (Shown at ~2B parameters; 4B available depending on hardware.) - **Desktop demo (LM Studio, M4 Mac, 48 GB unified memory):** running the 26B model (~26 GB in RAM including context) as an orchestrator farming translation sub-agents to multiple windows, translating the Gemma 4 announcement into many languages and compiling results into a web page — demonstrating local agentic and multilingual capability. ### Recommended adoption steps 1. Minimal code change: point an **OpenAI-compatible interface** at Ollama or LM Studio with a Gemma model. 2. Drop into **existing workflows** to see what works, what needs tuning, what's out of depth. 3. **Bolster evaluation suites** — "how good the model is depends on how well it does on your task, not anybody else's task"; benchmarks only indicate general capability. 4. Consider **serving realities**: self-hosting means controlling uptime but bearing maintenance, ongoing costs, and capex; mobile offloading requires thinking about device accelerators and RAM, but unlocks offline work and data that never leaves the device. --- ## Preserved Qualifications & Tensions - Gus explicitly states the open models are **not the most intelligent**, and admits his own bias. - Gus presents language fine-tuning difficulty as simultaneously **good and bad**. - Ian repeatedly frames claims as **task- and threshold-dependent** calculations rather than universal superiority of open models, and flags self-hosting's **maintenance and infrastructure costs** as real trade-offs. # Speaker Attribution & Provenance Record **Video:** Sovereign Escape Velocity: Ownership w Open Models — Gus Martins, & Ian Ballantyne, Google DeepMind **Video ID:** `01a15ca9-0c60-7d88-849f-439e338dff85` ## Speakers & Roles | Speaker | Role | Evidence | |---|---|---| | **Gus Martins** | **Co-presenter / creator-side speaker**, Google DeepMind, works on the Gemma product | Self-identified: *"I'm Gus... I work on the Gemma product."* Delivers first portion of talk. | | **Ian Ballantyne** | **Co-presenter / creator-side speaker**, Google DeepMind | Introduced by Gus: *"This is Ian. We are from Google DeepMind."* Delivers second portion and demos. | No guests, narrators, or other roles appear. Audience interaction is limited to informal polling (show of hands, verbal confirmation). ## Referenced Third-Party Sources & Claims - **LM Arena (referred to as "Lemarina" / "LM Marina")** — Elo scores and language rankings cited as benchmark evidence for Gemma 4 models. - **State of AI report by OpenRouter** — graph cited by Ian on token-generation volumes by task type; audience directed to the link shown on slide. - **OpenAI-compatible interfaces, Ollama, LM Studio** — cited as tools for running local models; LM Studio used in a live demo. - **Named third parties mentioned in passing:** Omar (keynote speaker, following day), CassD (follow-up talk, next day). ## Institutional / Sovereign-Use Claims (attributed to presenters, Google DeepMind) - Ukraine used Gemma in parts of government services. - Bulgarian LLM fine-tuned from Gemma 2; Brazilian Portuguese version fine-tuned from Gemma 3 (per Gus). - MedGemma variant for medical use cases (per Ian). ## Product & Promotional Segments - **Flagged as promotional:** Repeated endorsements of Google products throughout, including Gemini (proprietary) and the newly released **Gemma 4 family** (described as released "last Thursday"), the four model sizes (E2B, E4B, 26B, 31B), **AI.dev** (free trial venue), **Google AI Gallery app** (iOS/Android), and on-device demos. Statements such as *"we are very proud," "I strongly recommend you try those," "they are free"* are promotional in character. - Ian explicitly notes capability limits (*"I wouldn't get this model to do... a full systems architecture"*) — a scope statement by the presenter, not an external assessment. ## Provenance / Availability Notes - **Comment availability:** Not specified in supplied material — no transcript or metadata indicates whether comments are enabled. - **Demo limitations:** Gus states he *"was going to show a demo but I can't now"*; Ian's phone demo audio was muted, and one terminal demo required a retry. - Transcript contains verbatim errors (e.g., "Lemarina" for LM Arena, "coging aic") attributed to transcription, not to speakers' claims. Hi everyone. Uh, can you hear me? Yes, you can hear me. Hi. Sorry. Sorry we're 1 minute late. I'll try to do my best to finish earlier so my friend can do a pretty cool demos for you. I'm Gus. This is Ian. We are from the uh Google deep mind and I'm specifically I work on the Gemma product. Do any of you know what Gemma models are? Okay, perfect. Perfect. Thank you very much. Uh so today we're going to talk a little bit about ownership and open models. Uh and 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. Uh there's going to be more information tomorrow in the keynote by Omar and there's another talk by CassD also tomorrow that she go into even more details. We're going to tell a little bit of the story but the story is a little bit bigger. 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. Gemini is the best model we have. pretty strong multi model can do all kinds of things. But then there is uh there's more to this story than just having the strongest model possible. 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. 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. So when you think why does Google have two family of models? Because they complement each other. So Gemini is the most intelligent one. Can do a lot of cool stuff but it's hosted in Google servers. 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. 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. Uh two are target to mobile let's say like that or or IoT or smaller devices. It's a E2B and a E4B. Uh these names is a little bit weird. 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. The 2B is around 5B uh uh parameters. But then you say oh but where is these other 3B in memory the the fun fact is that they are not really parameters from the transformers. 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 2 billion or the 4 billion. 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. The E2B E4B both of them have a text uh vision and audio input and they do only text output. They can do thinking, they can do coding, function calling, all this kind of cool things. This all running on your phone right now. You could download it right now. Right. We also have other two models which are the larger ones. We have a 26B and a 31B. 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 where each one of those are like a 4B model. Why does it matter? Because it it has 26 billion parameters but it needs a space of a four billion parame to do the work. 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 31B dense which is 31 billion parameters model. And this is really really strong. If we look into our uh ELO score on LM Marina, you can see that both our models are they are I guess now they're fourth and seventh as the leads on open source models open models. And if you compare them to maybe the top 20 30 all of them are at least twice three times larger than our models in some cases 20 times larger. So we are talking about a disproportionate amount of intelligence per size. So our 31B model is the one I use very regularly. It can do basically anything from coging aic everything multilingual all of that. So I strongly recommend you try those. They are so strong that they are uh all both of them are really good to use on your on as a cloud deployed model. They can run on your desktop but if you use on your server as your endpoint to do uh your work they are pretty good. And you ask is this the most intelligent model? 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 mail, to do some more minial tasks, to help you code, to do some agentic capabilities that are searching and interacting with docs? 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 31B running one GPU. The competitors need 200 GB of memory which would be maybe four or five GPUs. So you can see that the price here is really really different. One easy place for you to try these models is on AI.dev where you can try Gemini models VO all the other mods but G gem are there both 26 and 31B 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 can't now but they can do some cool stuff which is vision plus thinking plus code execution all at the same time right uh I try to post something about this later but the ideas you can play with the models pretty easily and right there right not now finish let's finish the the talk and then you play with it uh and as I was saying the the intelligence per parameter that these models bring is pretty good it's very very strong and if We use the ALO score for Lemarina because it's a benchmark that's a person's preference, right? 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. So this is why this is so important. And 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 and sovereignty 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 you cannot use this model anymore it's all available to you and one of the changes we made last year until Gemma and others. We had our specific license, a Gemma license, which is pretty good, commercial friendly and all, 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, and then they will spend like 18 months doing procurement process to understand the license and trying to change. And that never works. So, it's pretty hard for sovereign institutions to adopt this kind of thing. That's why we move to a pass 2.0 you for for Gemma 4 and going forward and that makes thank you and that makes our life m your life much easier to to convince your legal department let's say like that that look we own this model we can use so this is pretty important and it enables many many uh sovereign institution to use our models we have some examples for example here for example uh Ukraine use Gemma to parts of their uh services we have 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 we also have a Brazilian version that is based on Gemma 3 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. And the problem is hard because not the the tooling or anything because the model is pretty strong on those languages already. So any gains you try to have you might not get there. So you might spend a lot of time to get like 1%. 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 because but it's good that you can automatically use in many language you can try right now and if you go into uh the marina for languages in many languages are top two three and look it's a 31B model. It's very very small right? So this is pretty good. That being said, uh I will let my colleague continue and show some demos. \>\> Thank you, guests. So, uh one thing that's I think is really important about these models is that when you think about using open models, you think about like using proprietary models. 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. 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. Uh this graph on the right hand side is from the state of AI report that open router did and it shows the uh it's a bit small for you on this diagram but have a look at that link. Uh shows the different types of tasks that people are doing through open router at the moment and you'll see that 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. 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 in itself. 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 Gus talked about doing some coding that's suitable for that then you have a GPU that you can take advantage to do some of that stuff. Now similarity 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 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 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 the right amount of latency depending on the that if it's a task for a user 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 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. So these are going to 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 can I fully offload or can I fully own compared to relying just on using the best possible models 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 going to run this little demo in the side at the moment. So we now have models that will work directly on mobile and edge devices. Uh this example here built by Coremax team. Um is a set of agent skills that the model is running on a phone. So uh I'm going to 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. 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. 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. So what this app will allow you to do is is kind of acts as like a playground. Uh so this is Google AI gallery and you can find it on iOS and Android 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 4 billion parameter model depending on uh the size of your hardware. And when we get to desktops and single GPUs, as Gus mentioned, that's where you can use the uh the 26 and the 31B models again on your local hardware. And I'll show you how to do that in a minute. But there's one kind of key point here is 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 costs. If we think about it, because now you're thinking about utilization of GPUs, you're thinking about utilization of MPUs on the hardware itself. Uh when are you going to do these tasks? Does the user need to get a response right now when you're taking a picture of something 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 ondevice or ownership because you think more about how they're 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 31 billion parameter model, you can now be thinking about scaling that down. So maybe if you wanted to use a 300 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. 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 trying to do. And the and the 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. So for instance, we have a variant of Gemma models called Medgema, 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 re feasibly deploy this to like one or maybe two GPUs uh to run that for I don't know like a whole hospital for instance. So these are kind of worth considering for the enterprise case. Uh I'm going to jump straight to demos now. Um I shown you some demos on the phone. I'm going to show you a quick demo uh here. Who quick show of hands. Who's ever used uh a tool called LM Studio? Okay, 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. And I'm at the moment, including the context, it's probably about 26 GB in RAM. And this is an M4 Mac. So, I've got unified memory. I've got up to about 48 gigabytes. Uh, so I can run it on this machine. And I'm just going to run this terminal right here. Let's give that a go. Oops. Pre-showing my demo. Let's try that again. There we go. So, I'm just going to run a little process where I'm going to do some quick a trick quick translation on my device. So, what it's going to do is I've got an orchestrator on this side here, uh, which is going to hopefully kick off my agent in a minute. 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 going to farm out uh, this translation to all of these different uh, uh, windows and each one of them represents a different sub aent. 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 want to want to translate to all these different languages. So you'll see in a second it should hopefully send it over there 3 two one and hopefully we should be generating translations a second. There we go. So so you can imagine doing any kind of aentic task on your local machine. uh you could have it like processing files, you could have it doing 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 web page and then you can see the results through the translation. There you go. So there's the multilinguality of the model there as well. Okay. Uh 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. Um you can take any open AI compatible interface that you've got and you can point it at a service like Olama or LM LM LM Studio and you can just pick out the gemma model and that's all you need to change codewise to at least uh try it out. 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? 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 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. So if you need to run your own GPU, you need to host it. Yes, you're in control of like uptown uptime and downtime, but then there's like maintenance costs and other stuff like that. 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. Um on mobile devices for instance you have to think about like if I'm going to offload stuff to a phone like what am I supporting? What accelerators do they have? What size RAM do they have? So the conversation becomes a little bit more complex but then there's a whole heap of things you can unlock like working offline or working on users private data that never leaves their device. Um and finally if you want to 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. 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. 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 want to help you on that journey as well. So, with that, thank you very much.