Conclusion

Let’s get straight to the point: the cost of running AI in-house has dropped below 5,000 yen a month.

What does this mean?

The era of thinking “AI is for large corporations,” “implementation costs millions,” and “it’s impossible without specialized personnel” is quietly coming to an end.

Liquid AI has announced a 230M (230 million) parameter model, along with a new fine-tuning method called “SlideFormer.” The reality that LLMs can run even on boards with only 8GB of memory dramatically lowers the “minimum cost” for SMEs to utilize AI.

The question is, “What changes as a result?”

—

What is the 230M Model?—The Truth Behind ‘Small but Usable’

Liquid AI’s 230M model has a mere 230 million parameters. Considering that GPT-4 is estimated to have over 1 trillion parameters and even the smallest version of Llama 3 has 8 billion parameters, this is two orders of magnitude smaller.

However, the important point here is that “small does not equal unusable.”

This model is designed to operate in environments with limited memory and CPU, such as smartphones, Raspberry Pi, and IoT devices. The “Liquid Foundation Models (LFM)” architecture developed by Liquid AI features a structure different from traditional Transformers, achieving high efficiency with fewer parameters.

What can it specifically do?

• Text Classification and Summarization: Automatic classification of inquiry emails, summarizing daily reports.
• Template Generation: Drafting texts for quotes, FAQ responses.
• Anomaly Detection in Sensor Data: Real-time assessment of temperature and vibration data on manufacturing lines.

If asked whether it can do the same things as GPT-4, the answer is no. However, 80% of the AI processing truly needed in the field of SMEs involves these “unassuming but reliably labor-saving” tasks. There’s no need for 1 trillion parameters for that.

—

SlideFormer—The Significance of Fine-Tuning with a Single GPU

To “adapt an AI model to your business,” fine-tuning is necessary. This has been the biggest bottleneck for SMEs until now.

Traditionally, fine-tuning a model with billions of parameters required multiple GPUs in the 1 million to 5 million yen range, such as NVIDIA A100 or H100. Even renting in the cloud costs several thousand yen per hour. A proper tuning could easily run into hundreds of thousands of yen.

Liquid AI’s SlideFormer flips this structure on its head.

Here’s how it works: instead of loading the entire model into GPU memory at once, it processes layers sequentially using a “sliding window” approach. While calculations are being performed on the GPU, the next layer is transferred from the CPU to the GPU, and the processed layers are sent back to the CPU. This pipeline processing reduces GPU/CPU memory usage by about 50% compared to traditional methods while maintaining over 95% peak performance.

What happens as a result?

With a single RTX 4090 (retail price 200,000 to 250,000 yen), you can fine-tune models with billions of parameters.

What used to cost “4 GPUs for 4 million yen” is now reduced to “1 GPU for 250,000 yen.” The cost is now 1/16th.

Even when using the cloud, if a single GPU instance suffices, the monthly cost is around 10,000 to 30,000 yen. If fine-tuning occurs about once a month, the effective cost can be kept to just a few thousand yen.

—

LLMs Running on an 8GB Board—The Collapse of ‘Inference Costs’

It’s not just fine-tuning. The costs of inference (the process of running AI to produce answers) are also collapsing.

For a model in the 230M class, single-board computers with 8GB of memory, such as the Raspberry Pi 5 (about 15,000 yen) or Jetson Orin Nano (about 30,000 yen), are sufficient to operate.

Electricity costs are just a few hundred yen a month. Even including hardware depreciation, the monthly cost is below 5,000 yen.

While there’s also the option of using cloud APIs, consider the value of “not having to send data outside.” For SMEs, there’s a significant resistance to sending quality data from manufacturing, personal information of customers, and sales know-how to external servers. If AI can operate entirely on their own board, that psychological barrier drops to zero.

For less than 5,000 yen a month, a dedicated AI can operate without sending company data outside. This is the reality for 2025.

—

So, Where Should SMEs Start?

Let’s move past the technical discussion and get into “what should we actually do?”

Step 1: Inventory Your Company’s ‘Repetitive Tasks’

Before implementing AI, first identify the “tasks currently done by humans that can be patterned.”

• Checking and transcribing order emails every morning
• Initial classification of inquiries
• Visual inspection of inspection images
• Aggregating daily and weekly reports

These are the types of tasks that involve “judgment but not high-level decision-making.” Determine how many hours are spent monthly and how much that equates to in hourly wages. Start by putting those numbers down.

Step 2: ‘Try It First’ with Existing APIs

There’s no need to immediately build a custom model. Use APIs from OpenAI or Claude to test whether the tasks identified in Step 1 can be automated. API costs often range from a few hundred to a few thousand yen a month.

Tasks that confirm “there’s a benefit” will become candidates for the next step.

Step 3: Fine-Tune with Your Own Data

For tasks that have shown effectiveness, if there are reasons such as “I don’t want to send my data outside,” “the API response speed doesn’t fit my business needs,” or “I want to stabilize monthly costs,” then proceed to build a custom model.

Base it on Liquid AI’s 230M model or Microsoft’s Phi-3 Mini (3.8 billion parameters) and fine-tune it using SlideFormer or LoRA (Low-Rank Adaptation). One workstation equipped with an RTX 4090 is sufficient.

Initial investment will be around 300,000 to 500,000 yen. If outsourced, it may be around 1 million yen. However, considering that traditional AI implementation costs between 5 million to 10 million yen, this is less than one-tenth.

Step 4: Set It Up to ‘Run Automatically’ at the Edge

Install the fine-tuned model on a Raspberry Pi or Jetson and set it up on-site. Next to the inspection line, behind the reception counter, or at the entrance of the warehouse.

No communication with the cloud, no latency, no monthly fees. Once set up, it will be in a state of “running automatically.” This is the strongest form for SMEs. It doesn’t become dependent on any individual. Even if the person in charge leaves, it won’t stop.

—

The True Competitive Advantage is ‘Being Able to Move Faster than Large Corporations’

For large corporations to implement AI, they must go through approvals, security reviews, vendor selection, PoC, and transition to production—taking at least six months, often a year.

For SMEs, if the CEO says “let’s do it,” they can be operational by next week.

Download the 230M model, load it onto a Raspberry Pi, and test it with your company’s inquiry data. Technically, this can be done in a single weekend. The only cost is the 15,000 yen for the board.

You may be tired of hearing the term “democratization of AI.” However, the world where a dedicated AI can operate for less than 5,000 yen a month is proven not by words but by numbers.

The question is simple.

How much does your company’s ‘repetitive tasks’ cost per month?

If that number exceeds 5,000 yen a month, there’s no reason not to start.