The Week in Tech: Open-Source AI Goes Mobile, EVs Break Range Records, and AI Datacenters Gobble Megapacks

Open-Source AI Goes Mobile: Gemma 4 12B

On June 3, Google DeepMind released Gemma 4 12B, a multimodal model built for agentic workflows on everyday hardware. At 12 billion parameters, it bridges the gap between Google's edge-friendly smaller models and its more powerful 26B mixture-of-experts variant—but the headline is its architecture. Unlike traditional multimodal models that rely on separate vision and audio encoders to translate inputs before the language model sees them, Gemma 4 12B eliminates those encoders entirely.

A Unified, Encoder-Free Architecture

Google replaced the vision encoder with a lightweight single-matrix embedding plus positional normalization, and it removed the audio encoder altogether, projecting raw audio signals directly into the same dimensional space as text tokens. The result is lower latency, reduced memory overhead, and a cleaner pipeline—all while supporting native audio input, a first for a mid-sized Gemma model.

Laptop-Ready Performance

Benchmark performance reportedly nears the larger 26B model, yet the footprint is small enough to run on a laptop with 16GB of VRAM or unified memory. For developers, that means state-of-the-art agentic capabilities without a cloud dependency. Google has published checkpoints on Hugging Face and Kaggle, and integration guides already exist for LM Studio, Ollama, llama.cpp, MLX, SGLang, vLLM, and Unsloth—meaning the ecosystem adoption will likely outpace many previous open-weight releases.

Open Weights, Open Ecosystem

Released under Apache 2.0, Gemma 4 12B also ships with Multi-Token Prediction drafters to reduce inference latency, and Google has launched an official Skills Repository to help agents build with the latest Gemma advancements. Cumulative Gemma 4 downloads have now crossed 150 million, with community projects ranging from wearable robotic arms to enterprise AI security.

AI's Hidden Energy Toll: $1B in Megapacks, Still Tied to Gas

While open-weight AI models run on laptops, the largest AI training clusters are anything but energy-efficient. SpaceX's amended S-1 filing, published June 4, reveals that xAI purchased $269 million in Tesla Megapack products in April 2026 alone—a single month that exceeds all of 2024. Total Megapack spending since 2024 now sits around $1 billion.

The Numbers Behind xAI's Tesla Orders

The filing breaks down related-party transaction granularity: $191 million in 2024, $506 million in 2025, $34 million in Q1 2026, then $269 million in April. Beyond Megapacks, SpaceX bought $131 million worth of Cybertrucks at MSRP in 2025, and Tesla recognized $573 million in revenue from SpaceX and xAI in 2025—roughly 4.5% of Tesla Energy's total revenue that year.

The Natural Gas Problem

Here is the tension Megapack purchases paper over: xAI's Memphis data center runs nearly 50 gas turbines, has won permits for a 41-turbine natural gas plant in Southaven, and could eventually emit more than 6 million tons of greenhouse gases per year. A planned 100 MW solar farm adjacent to the Colossus facility is a token gesture compared to the gigawatts of gas generation xAI is deploying. Battery storage is a critical grid-reliability tool, but when that storage is charged by burning gas, the clean-energy narrative collapses.

Summer EV Range Test: BMW iX3 Tops the Charts

Every summer, the Norwegian Automobile Federation (NAF) and Motor Magazine run a real-world EV range test that strips away WLTP marketing numbers. This year's winner was BMW's new iX3 50 xDrive, which covered 485 miles (781 km) on a single charge—slightly above its official WLTP rating and the longest result in the test.

The Full Leaderboard

Lucid Gravity placed second at 447 miles, followed by Mercedes-Benz CLA 350 4Matic (419 miles), Mercedes-Benz GLC 400 4Matic (413 miles), XPeng X9 (401 miles), Polestar 3 (373 miles), and a mix of Toyota, Kia, and Hyundai models rounding out the top ten. Several entrants beat their official WLTP figures, including XPeng X9 (+11.4%) and Kia EV2 (+5.4%), suggesting manufacturers are beginning to calibrate ratings more conservatively.

What the Test Actually Proves

The NAF test matters because it measures real-world highway efficiency, not a lab cycle. All cars start fully charged in Oslo, drive until they can no longer maintain the speed limit, and use normal mode with adaptive cruise control. BMW's 30% claimed improvement in range and 30% faster charging speeds appear credible after this test. With US orders starting at $61,500 and EPA-estimated 434 miles, the iX3 is now the benchmark for compact luxury EV range.

Foxconn Launches the Cavira: From iPhones to EVs

Foxconn finally unveiled a mainstream EV under its own Foxtron brand: the Cavira, a midsize crossover built to compete directly with the Tesla Model Y. It offers two trims—a rear-motor 186 kW long-range version with 359 miles of range, and a dual-motor 349 kW performance edition that hits 0-60 mph in 3.8 seconds.

Specs That Matter

The Cavira rides on a 2,920 mm wheelbase, uses an 82.7 kWh LFP battery, and charges from 10-80% in under 30 minutes at 175 kW. Inside, there is a 15.4-inch center touchscreen, a digital instrument cluster, and a 12-speaker sound system. Standard safety features include driver monitoring, blind spot detection, 360-degree cameras, and adaptive cruise control with lane following. It even supports V2L at up to 1,900 W for powering tools or appliances.

Can an iPhone Maker Disrupt the EV Market?

Pricing starts around $40,000, but the Cavira will almost certainly not see US sales. What makes it significant is Foxconn's manufacturing muscle: the company makes iPhones at scale, and it recently began producing its own LFP battery cells in Taiwan. If Foxtron can execute volume production reliably, it represents a credible contract-manufacturing threat to legacy automakers—and potentially a template for other consumer electronics giants entering the EV space.

Nissan Bets on Solid-State Batteries to Compete With China

Nissan has announced a three-year collaboration with Gelion and the University of Oxford to develop low-cost, resilient solid-state EV batteries. Dubbed Cost-effective, Resilient Solid-state Li-S, the project replaces expensive nickel and cobalt with sulfur, a material that is abundant and inexpensive to source.

The Sulfur Advantage

Gelion's Nano-Encapsulated Sulfur cathode platform swaps costly rare materials for sulfur, which can theoretically deliver cost advantages over Chinese-dominated lithium-ion supply chains. Longspur Capital Limited published a research report calling the technology Cheaper Than China, noting that Gelion's sulfur-based cathodes could be produced more cheaply in the West than comparable Chinese offerings today.

Timeline to 2028

The project aims for a commercial prototype by FY2027, aligning with Nissan's pledge to launch its first solid-state battery EV in 2028. Nissan opened its first all-solid-state battery production line at its Yokohama plant in January 2025 and is partnering with LiCAP Technologies for mass production using a dry-electrode process that eliminates solvent drying steps. While several Chinese brands, including BYD, plan limited solid-state deployments by 2027, Nissan's move signals a Western pushback in the battery technology race.

Wind and Solar Surpass Gas for the First Time

For the first time in history, wind and solar generated more electricity than gas globally in an entire month. According to independent energy think tank Ember, renewables produced 22% of the world's electricity in April 2026 versus 20% from gas—a record 531 terawatt-hours versus gas's 477 terawatt-hours.

Record Generation Numbers

Five years earlier, in April 2021, gas generation was nearly identical at 476 TWh, but wind and solar combined managed only 245 TWh—less than half of this April's output. Growth was essentially universal: China +14%, EU +13%, UK +35%, US +8%, Australia +17%, Chile +24%, and Brazil +4%. Ember's data showed no widespread reversion to coal despite energy-security concerns tied to Middle East tensions.

What It Means for the Energy Transition

The milestone is structural, not seasonal. April typically benefits from spring wind patterns and rising solar angles, but Ember's Global Electricity Review confirmed that wind and solar met all global electricity demand growth in 2025. Governments are accelerating targets accordingly: Indonesia plans 100 GW of solar-plus-storage, South Korea targets 100 GW of renewables by 2030, and the Philippines, Thailand, and UK are all fast-tracking deployment. The economics have flipped: renewables are now the cheapest, most secure form of new power in most markets.

UC Berkeley's AI Alarm: Failing CS Grades and Hollow Math Skills

Not every AI story is positive. UC Berkeley instructors are raising alarms after spring 2026 computer science courses saw failing rates far above department guidelines. In CS 10, 35.3% of students received F's—more than five times the 7% guideline. In CS 61A, 10.6% failed, also above the approved threshold. Both classes averaged C-plus, translating to a 2.3 GPA.

The Data Behind the Crisis

Teaching professor Dan Garcia identified two primary drivers: a vast increase in academic dishonesty fueled by LLMs like Claude, ChatGPT, and Gemini, and insufficient mathematical preparation. Nearly 30 students in CS 10 were caught cheating on take-home exams. But the problem extends beyond enforced dishonesty: students who leaned too heavily on LLMs for homework simply arrived at exams unprepared to solve problems independently. In EECS 127, an optimization course, 16.8% of students failed—far above the 5% typical for upper-division courses. Some students had taken linear algebra classes with open-internet, open-AI policies, leaving them unable to perform prerequisite-level proofs.

What Professors Are Doing About It

More than 1,300 UC faculty have petitioned to reinstate ACT and SAT standardized testing for STEM admissions to restore mathematical preparation standards. Both Garcia and Ranade are restructuring courses to emphasize critical thinking and resilience: clearer letter-grade thresholds, more low-stakes opportunities to practice, and explicit conversations about what confusion feels like in a learning environment. As Garcia put it, Confusion is the sweat of learning—and a growing number of students are skipping the workout.

What to Watch

These threads are not isolated. AI models are getting smaller, faster, and more open—but they are also demanding more power, creating a feedback loop between compute expansion and energy infrastructure. EVs are proving they can match or exceed combustion-vehicle range, but the battery supply chain remains a geopolitical battlefield. Renewables crossed a symbolic threshold, yet AI datacenter growth threatens to offset those gains unless the power sources themselves decarbonize. And in the classroom, the same AI tools transforming industry are eroding the foundational skills that industry claims to value. The weeks ahead will determine whether these signals turn into sustained trends or short-term noise.

Sources: Google DeepMind Gemma 4 12B announcement; SpaceX S-1/A filing (April 2026 Megapack disclosures); Norwegian Automobile Federation / Motor Magazine Summer EV Range Test 2026; Electrek coverage of Foxtron Cavira and Nissan-Gelion solid-state battery collaboration; Ember Global Electricity Review April 2026; UC Berkeley Daily Californian reporting on CS 10 and CS 61A grades.