“The Day the Belief that ‘Larger Models Equal Justice’ Ended”

For a long time in the world of AI, there was an unspoken rule that the more parameters a model has, the better it is. GPT-4 is estimated to have over a trillion parameters, and Gemini has several hundred billion. For companies wanting to use AI, servers equipped with multiple high-performance GPUs were necessary, requiring initial investments of several million yen. For small and medium-sized enterprises, it was a technology they could only observe.

That common understanding is quietly being overturned.

Mistral AI has announced its time series forecasting model, “Timestral.” With only 209M parameters (209 million), it surpasses existing models, including Google’s “TimesFM-2,” in zero-shot accuracy and has achieved top scores in major benchmarks such as GIFT-Eval, TIME, and fev-bench.

And here’s the most important fact: No GPU is required. It runs on a standard server’s CPU. The model’s weight file is only 837MB, small enough to fit on a USB memory stick.

“Cost Collapse” in Numbers

What exactly is changing? Let’s estimate the case of small and medium-sized enterprises introducing AI for demand forecasting and inventory optimization.

Using traditional large models:

• GPU-equipped server (equivalent to NVIDIA A100): Initial cost approximately 2 to 4 million yen
• Using cloud GPU instances (e.g., AWS p4d): Monthly cost 150,000 to 300,000 yen, annual cost 1.8 to 3.6 million yen
• Electricity and cooling costs: Annual cost 200,000 to 400,000 yen
• Total: Annual running cost approximately 2 to 4 million yen

Using the Timestral 209M model:

• Existing server or standard PC (CPU inference): Almost zero additional hardware investment
• Even when using the cloud, CPU instances suffice: Monthly cost 10,000 to 30,000 yen, annual cost 120,000 to 360,000 yen
• Electricity costs: Annual cost a few thousand to tens of thousands of yen
• Total: Annual running cost approximately 100,000 to 400,000 yen

At a minimum, this is one-fifth of the cost. Depending on the conditions, it could be less than one-tenth.

What used to cost 3 million yen a year is now 300,000 yen. This difference of 2.7 million yen is enough for a small business to hire one person. The phase shifts from “whether to use AI” to “what to do next with the money saved by using AI.”

Why Can Smaller Models Outperform Larger Ones?

This raises a question: How can a model with less than one-hundredth the number of parameters achieve better accuracy?

The key lies in the “trade-off between generality and specialization.” Large models like GPT-4 and Gemini are designed for versatility, capable of handling various tasks. Text generation, code generation, image understanding, translation—while they can do everything, they may be “over-equipped” for specific tasks.

On the other hand, Timestral is specialized for time series forecasting. It focuses on demand forecasting, sales forecasting, inventory fluctuations, and weather data analysis. By narrowing its scope, it achieves high accuracy with fewer parameters.

This is fundamentally important for small and medium-sized enterprises. On-site challenges are usually concentrated on specific tasks. They do not need an AI that can do everything; they want an AI that can accurately predict their inventory.

Why the World Bank Advocates for “Small AI”

In the same context, there is a noteworthy movement. Pascal Donohoe of the World Bank has proposed the “Small AI” concept.

In developing countries, many regions lack high-speed internet and stable electricity. The advanced country model of AI utilization, which relies on connecting to the cloud to access large models, simply does not work. This is why Donohoe argues for lightweight models that can operate on mobile devices or edge terminals.

AI that can predict agricultural harvest times, forecast infectious disease outbreaks, and optimize educational content—all operable on a single smartphone, even offline. The World Bank believes this could “shorten the speed of economic development by several generations.”

Is this discussion only relevant to developing countries?

The structure is the same in rural Japan. Areas with unstable high-speed connections, small and medium-sized enterprises that cannot invest in IT infrastructure, and sites without dedicated engineers. The conditions that make “Small AI” relevant apply equally to rural Japan.

The True Meaning of the “Large Model is Unnecessary” Argument

To clarify, it is not that large models will become unnecessary. General models like GPT-4 and Claude still excel in complex reasoning and creative tasks.

However, 80% of the AI required in business settings is actually for automating specific tasks.

• Sales forecasting
• Inventory optimization
• Anomaly detection
• Demand fluctuation analysis
• Generating standard reports

For these tasks, models with hundreds of billions of parameters are not needed. 209M is sufficient. In some cases, even 5M will do. Timestral offers four sizes: 5M, 15M, 58M, and 209M, allowing users to choose based on task complexity. With a 5M model, it’s now possible to run it even on a smartphone.

What Small and Medium-Sized Enterprises Should Do Starting Today

For those thinking, “This is an interesting story, but it doesn’t concern us,” here are three suggestions.

1. List what you want to forecast in your company
Sales, customer numbers, inventory turnover rates, raw material prices, personnel allocation—there should be something that would make things easier if you could predict it. That will be the first application for the small model.

2. Try it on existing PCs or servers
Timestral is publicly available with open weights. It can be used commercially under the Apache 2.0 license. There’s no need to buy expensive GPUs. Start by running it on a regular PC you have in-house. The fact that it works will change the atmosphere within the company.

3. Reassess your budget with the assumption that “GPU costs will be zero”
If your company has ever requested a quote for AI implementation, most of that estimate was likely for infrastructure costs. If those infrastructure costs decrease to one-tenth, what can you do with the saved budget? Allocate it for data preparation, use it for on-site training, or run another AI initiative. The options will suddenly expand.

The Structure Has Changed. If You’re Going to Move, Now Is the Time

When the cost structure of AI changes, the first to benefit will not be large corporations. It will be the “groups that have been unable to use AI due to costs.” In other words, small and medium-sized enterprises.

3 million yen becomes 300,000 yen. No GPUs are needed. It can run on smartphones. This is not just a story about cost reduction. It represents a structural change where the boundary between those who can use AI and those who cannot disappears.

While large corporations invest billions in large models, small and medium-sized enterprises will acquire equal or greater predictive accuracy with the 209M small model. Companies that have recognized this reversal are already starting to take action.

While you say, “It’s still too early,” the company next door is already moving. When technology becomes cheaper, the only thing that will differentiate companies is speed.