Gemma 4 12B: Free Local AI Model for Your Laptop

Samantha: Welcome to The State of Tech, The European Edition, Saturday June six, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Meta puts facial recognition into smart glasses, Alphabet raises eighty billion dollars for AI infrastructure, Google DeepMind releases a laptop-friendly AI model called Gemma 4, IBM bets ten billion dollars on quantum computing, and to close out, two things you can actually try yourself: the open-source AI model Gemma 4 12B running locally on your laptop, and Amazon's natural-language warehouse robot Proteus heading to Europe. Let's start with Meta.

Meta quietly rolls facial recognition into smart glasses, sparking a fresh European privacy storm.

Samantha: Meta is reportedly pushing facial recognition code to millions of phones as part of its smart glasses platform. The idea is that the glasses can identify people you look at in public, using biometric data stored locally on the device.

Bob: And that local storage is the bit Meta will lean on when the questions come. The pitch is: your face data stays on your phone, not in a cloud database. But the practical effect is the same. You walk down a street in Berlin or Amsterdam, someone wearing these glasses looks at you, and a name could pop up in their field of view.

Samantha: This is the moment a lot of privacy lawyers have been warning about for years. Always-on cameras on your face, combined with software that recognises individuals, turns every wearer into a walking identification system.

Bob: And it lands directly in GDPR territory. Biometric data is treated as a special category under European law, which means stricter rules on consent and purpose. The person being identified hasn't agreed to anything. They're just walking past.

Samantha: Workplaces are another flashpoint. Imagine a colleague wearing these in an open-plan office, or a manager in a meeting room. Suddenly HR policies on recording and surveillance need a full rewrite.

Bob: Regulators in Ireland and France have already shown they're willing to take on Meta over data practices. The Irish Data Protection Commission has been the lead supervisor for Meta in Europe, and previous facial recognition rollouts by other companies were forced back. Clearview AI, for instance, was fined repeatedly across European countries.

Samantha: The technology question is one thing. The social question is bigger. Do we want a world where strangers can name you on sight? Polling across Europe has consistently shown low public appetite for that, even when the use case sounds helpful, like remembering names at a conference.

Bob: Meta hasn't formally launched the feature with a press event. The reporting describes the code arriving quietly on devices, which itself raises the question of transparency. If a feature is significant enough to change how public spaces work, regulators tend to want to hear about it in advance, not discover it from reverse engineering.

Samantha: The bottom line: this is a test of whether European biometric rules actually constrain the largest platforms, or whether enforcement always arrives years after deployment. Expect formal questions from Brussels and Dublin within weeks.

Alphabet raises eighty billion dollars to fund a massive AI infrastructure build-out.

Bob: Alphabet, Google's parent, is going to the market for eighty billion dollars through a stock offering. The money is earmarked for AI infrastructure: compute capacity, data centres, the physical backbone that runs models like Gemini.

Samantha: The framing here is important. Alphabet is essentially telling investors that the AI race has shifted. It's no longer mainly about who has the cleverest model. It's about who can build the biggest, most efficient computational network to run those models for billions of users.

Bob: And demand is outrunning supply. Google has been telling customers and developers for months that capacity is tight. Eighty billion is the answer to that bottleneck.

Samantha: For Europe, this matters in two ways. First, more capacity globally could ease waiting lists and bring down the cost of AI services for European businesses. Second, where Alphabet chooses to build will shape the European digital landscape for years.

Bob: Google already has significant data centre presence in the Netherlands, Belgium, Ireland, and Finland. Any expansion there brings jobs and tax revenue, but also energy and water demand, which has become a politically sensitive topic in several of those countries.

Samantha: The wider market context is that this raise puts Alphabet shoulder to shoulder with the other hyperscalers in terms of capital intensity. Microsoft, Amazon, and Meta are all spending at similar scale. European cloud providers simply cannot match these numbers.

Bob: Which loops back to the sovereignty debate. The more dependent European businesses become on American AI infrastructure, the louder the calls in Brussels and Paris for a European alternative. So far those calls have not translated into spending anywhere near this magnitude.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Google DeepMind releases Gemma 4 12B, a multimodal AI model that runs on ordinary laptops.

Bob: Google DeepMind has published Gemma 4 12B, an open-source AI model that handles text, images, and video, and is designed to run on a normal laptop rather than in a data centre.

Samantha: It's released under the Apache 2.0 licence, which is the permissive end of open source. Companies can take it, modify it, and deploy it commercially without paying Google or worrying about restrictive terms.

Bob: The technical novelty is that it processes images and video directly, without a separate encoder bolted on the side. The practical result is that it runs efficiently on limited hardware. You don't need a server farm to use it.

Samantha: For European businesses, this is significant. Running AI locally means sensitive data, customer records, medical files, legal documents, never leaves the device. That sidesteps a lot of the GDPR friction that comes with sending data to American cloud AI services.

Bob: It also opens the door for smaller European companies and public sector bodies that can't afford ongoing API bills. A model you download once and run forever changes the economics.

Samantha: The wider trend here is on-device AI catching up with cloud AI. Two years ago, anything useful required a connection to a remote server. Now we're seeing capable models small enough to fit on a laptop, and even on phones.

Bob: There's a strategic angle for Google too. By making Gemma open, they grow a developer ecosystem around their AI approach, even when they're not collecting subscription revenue directly. It's the same playbook Meta has used with its Llama models.

Samantha: For developers listening, this is one to download today. We'll come back to that in our closing two.

IBM commits over ten billion dollars to deliver a fault-tolerant quantum computer by 2029.

Bob: IBM is putting more than ten billion dollars over five years into quantum computing, with the goal of building what it calls the first large-scale, fault-tolerant quantum machine by 2029. The system has a name: IBM Quantum Starling.

Samantha: Fault-tolerant is the key phrase. Today's quantum computers make a lot of errors. They can do impressive things in lab demonstrations, but they're not reliable enough for serious industrial work. Fault tolerance means the machine can correct its own mistakes as it runs.

Bob: IBM claims this machine will execute twenty thousand times more operations than current systems. If that materialises, it opens up problems that are genuinely impossible on today's computers, things like simulating new materials, modelling complex molecules for drug discovery, or optimising logistics at continental scale.

Samantha: For Europe, the timing is interesting. The EU has its own Quantum Flagship programme, and several member states, Germany, France, the Netherlands, have national quantum strategies. None at this spending level, though.

Bob: IBM also runs a network of partners and research access programmes. Universities and companies in Europe have been using IBM quantum systems through the cloud for years. A working fault-tolerant machine would extend that access to far more serious workloads.

Samantha: One sober note: 2029 is a target, not a guarantee. Quantum timelines have historically slipped. But ten billion dollars in committed spending changes the credibility of the roadmap significantly.

Bob: There's also the cryptography angle. A fault-tolerant quantum computer could eventually break the encryption that protects banking, government communications, and most internet traffic. European regulators have been pushing for post-quantum cryptography standards, and this kind of timeline puts pressure on that transition.

Try Gemma 4 12B yourself, a free multimodal AI model running entirely on your laptop.

Samantha: To close out, two things you can actually try this weekend. First, the model we just mentioned: Gemma 4 12B from Google DeepMind.

Bob: It's free, it's open-source under Apache 2.0, and it runs on a normal laptop, no cloud account required. You can download it from Google's developer site or through the usual model hubs like Hugging Face.

Samantha: To run it comfortably, you'll want a machine with a decent amount of memory, sixteen gigabytes or more, and ideally a recent processor or a discrete graphics card. A modern MacBook handles it. A reasonably recent Windows or Linux laptop with a gaming-grade graphics card handles it.

Bob: What can you actually do with it? Summarise long documents without sending them anywhere. Describe what's in a photo. Pull text out of a screenshot. Ask it questions about a video file. All offline, all on your machine.

Samantha: For anyone who's been nervous about pasting sensitive material into ChatGPT or Gemini, this is the alternative. The data simply never leaves your laptop.

Bob: The easiest way in for non-developers is a tool called Ollama, which gives you a simple interface to download and chat with models like Gemma. Install Ollama, pull the Gemma model, start chatting. Fifteen minutes from zero to running.

Samantha: For developers, the model is fully accessible through standard libraries, and the Apache licence means you can build commercial products on top without paying royalties.

Bob: One caveat: a 12 billion parameter model is capable, but it's not GPT-5. Expect it to be useful for everyday tasks, not for the absolute hardest reasoning problems. For most people, that trade-off is well worth the privacy.

Samantha: A practical weekend project: try summarising a stack of PDFs you've been meaning to read, all locally.

Amazon's Proteus warehouse robot now takes plain-English instructions and is coming to Europe.

Bob: The second thing worth flagging is Amazon's updated Proteus robot. It's part of an eleven point six billion dollar investment in European fulfillment, and it's scheduled for European warehouses in the first half of 2027.

Samantha: The new bit is that warehouse workers can now instruct Proteus using natural-language text prompts. You type what you want in plain English, or presumably French, German, or Spanish in time, and the robot figures out the priorities and the route on its own.

Bob: That's a meaningful change. Previous warehouse robots needed specialised programming or rigid command sequences. Now an employee with no technical training can redirect a robot mid-shift.

Samantha: Proteus also now moves across entire warehouse floors, not just the loading dock areas it was previously limited to. So it's a much bigger footprint inside the building.

Bob: For European workers, the picture is mixed. On one hand, heavy lifting and repetitive movement get offloaded to machines, which has health benefits. On the other, the long-term question of how many human roles remain in a fully automated warehouse is real.

Samantha: Amazon's framing is that robots take on the physically demanding parts and humans focus on judgement and exception handling. Unions across Europe, particularly in Germany and Italy, have been pushing back on that narrative and want firm commitments on staffing levels.

Bob: You can't download Proteus, obviously, but if you order from Amazon in Europe in the second half of next year, there's a decent chance your package will have been moved by one of these.

Samantha: That's the day. Samantha, take us out.

Samantha: Today we covered: Meta's facial recognition smart glasses, Alphabet's eighty billion dollar AI infrastructure raise, Google DeepMind's open-source Gemma 4 12B model, IBM's ten billion dollar quantum bet, and two things to try yourself: Gemma 4 12B running locally through Ollama, and Amazon's Proteus robot coming to European warehouses next year.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.