Open Models Are Getting Smaller, Smarter, and More Practical

The open source AI movement continues to accelerate. Microsoft, Meta, Mistral, OpenAI, DeepSeek, and Google have all released open models in recent months. Google's latest Gemma 3 variant stands out for an unusual reason: it's just 270 million parameters, small enough to run directly on your phone.

To put that in perspective, many smaller AI models measure their parameters in billions—GPT-4 has over a trillion, Claude 3 has hundreds of billions. Parameters are essentially the learned connections that give AI its intelligence, so 270 million is extraordinarily small by today's standards. This tiny model running on your phone probably matches the capabilities of GPT-3.5 from just 18-24 months ago. The same intelligence that once required massive data centers now fits in your pocket.

For associations, this compression of intelligence is an important trend line to watch. While the biggest models tend to receive more press, the real revolution may be happening at the opposite end of the spectrum.

The Compression Timeline 

AI is getting cheaper and smaller at an exponential rate. What required billions of parameters two years ago now works with millions. Models are shrinking by orders of magnitude while maintaining similar capabilities. Gemma 3's tiny variant operates roughly 100 times faster than a 20-billion parameter model and two orders of magnitude faster than models in the hundreds of billions range.

We're watching the democratization of AI happen in real-time. State-of-the-art intelligence from 2023 is now accessible on the cheapest possible hardware. By the time your association finishes debating how to adopt AI, the technology will have compressed another order of magnitude.

The trend line matters more than any single model release. While tech giants compete over who has the smartest AI, the real race is toward who can deliver useful intelligence at the lowest cost and smallest footprint.

The Skip AI Philosophy: Breaking Big Problems Into Small Pieces

The power of small models becomes apparent when you rethink how to solve problems, and we've seen this firsthand in one of our sister companies. Skip AI, an analytics agent designed for enterprise data analysis, demonstrates this perfectly. Rather than throwing one massive model at complex queries, Skip breaks them down into dozens, sometimes hundreds of smaller tasks.

Let's say you need a comprehensive membership dashboard with 20-30 different components—retention trends, engagement metrics, demographic breakdowns, revenue analysis. It simply wouldn't be feasible to generate everything in one massive prompt to an expensive frontier model. Skip takes a different path.

It analyzes the request like a human team would, breaking it into component parts. Each piece—a retention chart here, an engagement metric there—gets assigned to a small model instance. These models run in parallel, generating first drafts. Slightly larger models test the code. Another layer composes everything together. The result: potentially hundreds of prompts executed efficiently rather than one expensive, slow, monolithic process.

You don't need hundreds of GPT-5 API calls for a single dashboard. But hundreds of tiny model calls? That's not only affordable—it's faster and often more reliable. When small models fail, they fail small. When large models fail, your entire operation stops.

The 80/90% Solution Works Better Than You Think

Many tasks don't require 100% accuracy. They require 80-90% accuracy delivered quickly and affordably.

Not everything requires perfection. Some things absolutely do—you definitely want 100% accuracy on an airplane ride, from takeoff through landing. No compromises. But AI for associations is more like hiring an assistant. Would you rather have an assistant who's right 100% of the time but costs $100,000 per month and takes hours to respond? Or one who's right 85% of the time, costs a tiny portion of that, and responds instantly?

Many member service inquiries don't need frontier intelligence. Basic content generation works fine with smaller models. Event registration confirmations, renewal reminders, initial member inquiries—these represent the meaningful AI opportunities, and they work beautifully with models a fraction of the size of GPT-5.

Using the most powerful model for everything is like using a sledgehammer to hang picture frames. Save the expensive, powerful models for the truly complex tasks. Use small, efficient models for everything else. Your members won't notice the difference, but your budget will.

What Changes When AI Approaches Free

When models shrink to 270 million parameters, they can run directly on phones, browsers, even smartwatches. No API calls. No cloud infrastructure. No usage limits. The marginal cost of inference approaches zero. More importantly, these tiny models require a fraction of the energy—we're talking about the difference between needing a data center's worth of cooling versus running on your phone's battery for hours. Google's testing showed Gemma 3 270M used just 0.75% of battery for 25 conversations on a Pixel phone.

This changes the entire calculus of AI adoption. Instead of carefully rationing API calls to manage costs, you can theoretically embed AI everywhere. Every member interaction, every piece of content, every analytical query can be AI-enhanced without checking the budget first.

Privacy concerns also evaporate when data never leaves the device. A member's sensitive information stays on their phone while still getting AI-powered assistance. Smaller associations without IT infrastructure can deliver AI capabilities that previously required enterprise-grade systems.

We are often impressed by what AI can do, but we should also be impressed where it can run. When intelligence fits everywhere and costs almost nothing, you stop thinking about "AI projects" and start thinking about "AI as default." Every form gets smarter. Every search gets more intuitive. Every interaction becomes more personalized.

What This Means Today

The small model revolution demands a shift in how associations approach AI strategy. Stop waiting for the perfect model—by the time GPT-6 becomes affordable and stable, models like Gemma will be running on smartwatches with similar capabilities.

Build with flexibility as your priority. Today's small model is tomorrow's tiny model. Design systems that can swap models like changing batteries. Focus on workflows and use cases, not specific model capabilities. The associations winning with AI aren't the ones with access to the most powerful models—they're the ones who've figured out how to use appropriate models for appropriate tasks.

Consider building a fleet of specialized models rather than relying on one general-purpose giant. Fine-tune small models for specific tasks: one for member inquiries and another for event management. Specialized small models often outperform general large models on their specific tasks while costing a fraction to run.

The future of AI for associations isn't one perfect model to rule them all. It's thousands of tiny, specialized models running everywhere, all the time, for basically free. Associations that understand this shift can AI-enable every touchpoint, process, and interaction—not just flagship features.

While everyone's looking up at the AI giants battling for supremacy, the real revolution is happening at the microscopic level. Your phone already has more accessible AI power than a data center from five years ago. The small model revolution is playing out right in front of our eyes. How will you take advantage of it?