Sovereign Escape Velocity: Ownership w Open Models — Gus Martins, & Ian Ballantyne, Google DeepMind
Google DeepMind's Ian Ballantyne and Gus Martins introduce Gemma 4, a family of open models delivering state-of-the-art performance with remarkable size efficiency. They discuss how models like the 31B variant outperform competitors 2-20x its size while running on a single GPU, the shift to an Apache 2.0 license to foster sovereignty and adoption, and the new economics of running powerful agentic workloads on hardware ranging from a Pixel phone to a single enterprise GPU.
Gus Martins and Ian Ballantyne from Google DeepMind introduce Gemma 4, a new family of open models designed to complement the proprietary, high-end Gemini models. While Gemini offers peak intelligence via API, Gemma provides ownership, control, and the ability to run on-premise, which is crucial for applications involving proprietary data or requiring deep customization.
The Gemma 4 Model Family
The launch includes four distinct models tailored for different hardware targets:
- E2B & E4B (Effective Models): These are designed for mobile and edge devices. The ‘E’ stands for ’effective,’ signifying their memory efficiency. For instance, the E2B model has approximately 5 billion parameters but is engineered to only require the GPU memory of a 2-billion-parameter model. This is achieved by offloading non-transformer parameters (like token embeddings) to standard memory, enabling them to run powerful multimodal (vision, audio, text input) and agentic workloads (coding, function calling) directly on a smartphone.
- 26B (Mixture of Experts - MoE): This model contains 26 billion total parameters but only activates about 4 billion for any given task. This MoE architecture makes it as accessible as a much smaller model, dramatically lowering the hardware barrier for deployment while retaining strong performance.
- 31B (Dense Model): This is the flagship and most powerful model in the Gemma 4 family. It delivers exceptional performance across a wide range of tasks, including coding, multilingual applications, and agentic reasoning.
Unprecedented Size-to-Performance Ratio
A key theme is the “disproportionate amount of intelligence per size.” Gemma 4 models consistently rank near the top of open model leaderboards like the LM Arena, placing 4th and 7th while being significantly smaller—in some cases, 2x to 20x smaller—than their direct competitors. This efficiency has profound economic implications:
- Hardware Cost: The 31B model runs on a single GPU, whereas competitors at a similar quality level might require four or five GPUs (around 200GB of VRAM).
- Accessibility: The models can be easily experimented with on platforms like
AI.devfor free.
Fostering Ownership and Sovereignty
To simplify adoption, Gemma 4 has moved from a custom license to the widely accepted Apache 2.0 license. This change is critical for sovereign institutions (e.g., governments, national labs) that often face 18+ month procurement and legal review cycles for custom licenses. This move has already enabled national-level AI initiatives in countries like Ukraine, Bulgaria, and Brazil to build upon Gemma.
Interestingly, the base models are now so proficient in multiple languages that fine-tuning for a specific language yields only marginal gains, a testament to their strong out-of-the-box multilingual capabilities.
New Economics for Agentic Workloads
As AI shifts towards more complex, agentic workflows, the cost of token generation becomes a significant factor. By owning and running models locally or on-premise, the economic calculation shifts from per-token API fees to hardware utilization and energy costs. This allows for extensive use in high-token tasks like coding, research, and data analysis without escalating API bills.
Ian Ballantyne demonstrated this with several compelling examples:
- On-Device Agents: An agent running on a phone using the E-series models to reason about and execute on-device skills, like interacting with calendar or map applications.
- Local Multi-Agent Systems: Using LM Studio on an M4 Mac, a 26B model was orchestrated to run multiple parallel sub-agents, each translating a document into a different language simultaneously.
- Enterprise Scale: A single H100 or even an L4 GPU can now run a model powerful enough for serious enterprise tasks. For example, a specialized medical variant like MedGemma could be deployed on one or two GPUs to serve an entire hospital, keeping all data private.
Getting Started
Integrating Gemma 4 is straightforward. You can use any OpenAI-compatible client, such as those provided by Ollama or LM Studio, and simply point it to a Gemma model. The speakers encourage users to drop Gemma into existing workflows to evaluate its performance on their specific tasks and to bolster their internal evaluation suites beyond generic benchmarks.