Niềm tin vào LLM sụp đổ: Mô hình nhỏ lên ngôi, Apple bất ngờ được lợi
Sự phát triển chậm lại của các mô hình ngôn ngữ lớn (LLM) như GPT-5 đang gây ra sự thất vọng, tương tự như những nâng cấp "tẻ nhạt" trên điện thoại thông minh.
Nhiều doanh nghiệp đang chuyển sang sử dụng các mô hình ngôn ngữ nhỏ (SLM), nhờ khả năng tùy chỉnh, chi phí rẻ và hiệu quả cao hơn trong các tác vụ cụ thể.
SLM có thể chạy trên hệ thống nội bộ của doanh nghiệp thay vì phụ thuộc vào nền tảng đám mây, đặc biệt phù hợp với các thiết bị như điện thoại, ô tô tự lái và robot nhờ tiết kiệm năng lượng.
Mô hình Nvidia Nemotron Nano (9 tỷ tham số) đã vượt qua Llama của Meta (gấp 40 lần kích thước) trong nhiều bài kiểm tra.
Gartner dự báo nhu cầu doanh nghiệp cho các mô hình chuyên biệt sẽ tăng gấp đôi tốc độ so với LLM trong năm 2025.
Ví dụ điển hình: IBM dùng mô hình chỉ 250 triệu tham số cho sản phẩm Docling, chuyển đổi dữ liệu từ PDF – tiết kiệm chi phí và không cần LLM.
Các SLM có thể hoạt động trên CPU thông thường thay vì GPU đắt đỏ – “GPU là siêu xe Ferrari khó chiều,” theo IBM.
Một nghiên cứu của Nvidia cho thấy SLM là tương lai của agentic AI, giúp tạo ra cấu trúc "giống Lego" với các mô hình chuyên biệt cho từng tác vụ, thay vì dựa vào một LLM khổng lồ duy nhất.
GPT-5 cũng có cấu trúc lai – sử dụng nhiều mô hình nội bộ với kích cỡ và khả năng khác nhau tùy tác vụ.
Apple được cho là đang tận dụng SLM để tăng sức mạnh cho "Apple Intelligence" trên iPhone, xử lý một số tác vụ trên thiết bị thay vì đám mây – cách tiếp cận này được đánh giá là thông minh và tiết kiệm.
Trong khi các "đại gia đám mây" như Microsoft và Google vẫn đầu tư mạnh vào LLM và hạ tầng siêu máy chủ, giới phân tích cảnh báo đó có thể là hướng đi thiếu linh hoạt nếu SLM tiếp tục phát triển nhanh.
📌 Khi LLM đang giảm dần hiệu quả và chi phí trở thành yếu tố then chốt, mô hình ngôn ngữ nhỏ (SLM) chứng minh sức mạnh nhờ chi phí thấp hơn gấp 10–30 lần và hiệu suất vượt trội trong các tác vụ cụ thể. Apple, với chiến lược chậm mà chắc và khai thác SLM trong "Apple Intelligence", có thể trở thành người hưởng lợi dài hạn trong cuộc đua AI.
WHEN TECH folk talk about the lacklustre progress of large language models (LLMs), they often draw an analogy with smartphones. The early days of OpenAI’s ChatGPT were as revolutionary as the launch of Apple’s iPhone in 2007. But advances on the frontier of artificial intelligence (AI) have started to look akin to ho-hum phone upgrades rather than genuine breakthroughs. GPT-5, OpenAI’s latest model, is a case in point. It has generated even less buzz than is expected from the iPhone 17, Apple’s newest release, which is due to be unveiled on September 9th.
The slowing pace of improvement at the bleeding edge of generative AI is one sign that LLMs are not living up to their hype. Arguably a more important indication is the rise of smaller, nimbler alternatives, which are finding favour in the corporate world. Many firms prefer bespoke models that they can tailor to their specific needs. These so-called small language models (SLMs) are cheaper than all-purpose LLMs, whose God-like intelligence can appear superfluous. As David Cox, head of research on AI models at IBM, a tech company, puts it: “Your HR chatbot doesn’t need to know advanced physics.”
Besides being as easy to run on a company’s in-house IT systems as via a cloud-service provider, SLMs may also be more useful for AI agents, which do work-related tasks alongside or instead of humans. SLMs’ slimmer size makes them particularly well suited for AI in smartphones, self-driving cars, robots and other devices where energy efficiency and speed are at a premium. If they continue to become more reliable, they could validate the decision of Apple and other device-makers not to follow the herd by investing gazillions in cloud-based LLMs.
Small is beautiful
There is no precise definition of what constitutes small versus large language models. The distinction comes down to the number of parameters they are trained on—ie, the amount of numerical settings in a model’s brain that help it make sense of data. LLMs stretch to hundreds of billions of parameters. SLMs may be trained on 40bn or fewer, down to fewer than 1bn in the case of tiny ones.
Improvements in the training of SLMs have helped them catch up with the capabilities of LLMs. Small models are increasingly “taught” by larger ones, rather than having to crawl the web to learn for themselves. Artificial Analysis, a benchmarking firm, says that on a variety of tests, a 9bn-parameter model called Nvidia Nemotron Nano, recently launched by the AI chipmaker, outperforms a Llama model released in April by Meta, a social-media giant, which is 40 times bigger (see chart). Catch-up is becoming routine. “The smaller models of today are much more capable than the larger models of last year,” says Mohit Agrawal of Counterpoint, a tech market-research firm.
Chart: The Economist
Better performance has attracted business customers. Gartner, an IT forecaster, says well-known flaws in LLMs, such as hallucinations, have bred “user fatigue”. Instead, businesses want more specialised models, fine-tuned on industry-specific data. This year, corporate demand for them is projected to grow twice as fast as it is for LLMs, though from a much lower base. Over time, Gartner expects more of these specialised models to be developed by companies in-house.
Another reason for the growing popularity of SLMs is economics. Companies have moved on from a spend-whatever-it-takes approach, employed in the early days of generative AI, to a greater focus on return on investment. Though they may still use LLMs for many tasks, they can save money by doing discrete, repeatable jobs with SLMs. As one venture-capital grandee puts it, you may need a Boeing 777 to fly from San Francisco to Beijing, but not from San Francisco to Los Angeles. “Taking the heaviest-duty models and applying them to all problems doesn’t make sense,” he says.
To illustrate the economics, IBM’s Mr Cox points to a product developed by his company, called Docling, which transforms PDF documents, such as receipts, into storable data. It runs on a “tiny” model with about 250m parameters. He calls it a useful tool, but one that would not be cost-efficient if it were run on an LLM. Small models can also run on cheaper varieties of chip. The smallest models can use central processing units (CPUs), the workhorses of general computing, rather than graphics processing units (gpus) like those that have made Nvidia the world’s most valuable company. This could be a big selling point: GPUs, Mr Cox says, are “fussy Ferraris that are always in the shop”.
SLMs may become still more attractive as businesses deploy more AI agents. A little-noticed paper published in June by Nvidia Research, the chipmaker’s investigations arm, states boldly that “small, rather than large, language models are the future of agentic AI.” It notes that currently most agents are powered by LLMs, hosted by cloud service providers. The investment pouring into AI-related cloud infrastructure suggests that the market assumes that LLMs will remain the engines of agentic AI. The paper challenges that assumption, arguing that SLMs are sufficiently powerful to handle agentic tasks, and more economical (for instance, a 7bn-parameter model can be ten to 30 times cheaper to run than a model up to 25 times bigger). It says that SLMs may lead to a “Lego-like” approach to building agents, with firms using small, specialised experts, rather than one monolithic LLM intelligence.
The paper, which raises questions about the money pouring into cloud-based agentic infrastructure, does not represent Nvidia’s strategic thinking, says Kari Briski, a senior Nvidia executive. She says business customers want models “of all shapes and sizes”. The largest models are better at handling the hardest jobs than small ones. Moreover, she says, it is important to keep pushing the frontier of LLMs because that makes them better teachers of SLMs.
Whether SLMs eventually displace LLMs or not, heterogeneity is on the rise. All-knowing LLMs will remain important for consumer apps like OpenAI’s ChatGPT. Yet even OpenAI is shifting its focus. GPT-5 has internal models of different sizes and strengths, which it uses depending on the complexity of the task.
As SLMs become more adept, they may also improve the reputation of on-device AI. Counterpoint’s Mr Agrawal expects Apple to play up the potential of smaller models as part of the launch of a slimmed-down iPhone on Tuesday. “Apple Intelligence”, the company’s AI product, which uses SLMs to do some tasks on the iPhone while delegating harder jobs to the cloud, could be the shape of things to come, he believes. “It’s not a case of missing the boat,” he says of Apple’s disappointing AI efforts so far. Given the firm’s clout, “it can always get on the next boat.”
For now, most attention remains on LLMs. Though cloud giants like Microsoft and Google have built extremely small models, the belief that the status quo will persist has helped justify the fortunes they are spending on data centres to train and host the largest models. That may be short-sighted. Given the relative merits of SLMs, Apple’s take-it-slow approach may turn out to be justified in the long run. ■
