news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Zhidongxi (public account:zhidxcom）
Compiled by Meng Qiang
Edited by Yunpeng

According to Venture Beat, Hugging Face, Mistral AI and OpenAI launched their own small models (SLM) on July 16 and 18 respectively, promising to popularize advanced natural language processing capabilities. In the past, technology companies competed in the pursuit of larger and more complex neural networks in the large language model track. Venture Beat believes that these small models have opened up new tracks and may also affect the way companies use AI solutions.

Small models, as the name implies, are relative to large language models (LLMs). They generally have fewer parameters and lower computing resource requirements. Compared with large language models with parameters ranging from hundreds of billions to trillions, the three new small models: SmolLM, Mistral NeMo, and GPT-4o mini can have parameters ranging from hundreds of millions to tens of billions, and are lower than large language models in terms of training volume and energy consumption. Although the three models use different methods to achieve AI popularization, they all have a common goal: to bring powerful language processing capabilities to more devices and applications.

How small models change edge computing

Venture Beat believes that Hugging Face's SmolLM is the most innovative of the three. It is designed to run on mobile devices and has three specifications: 135 million, 360 million, and 1.7 billion parameters, which facilitates AI processing on edge devices and solves key issues of data privacy and latency.

The significance of SmolLM goes far beyond improving efficiency. Enabling edge devices to use AI processing enables device applications to run on the basis of low latency and high privacy protection. In the past, many complex AI functions could not be realized due to privacy or connection issues. With SmolLM, these functions may become a reality.

In addition, Mistral AI launched the Mistral NeMo model, which has 12 billion parameters and a context window of up to 128k. Mistral NeMo is aimed at desktop computers and is positioned between large cloud models and ultra-compact mobile AI. The former is a large-scale AI model trained and run on a cloud computing platform, and the latter is an efficient and compact AI system running on resource-constrained mobile devices (such as mobile phones and wearable devices).

Venture Beat said Mistral NeMo's computing method may bring significant changes to the enterprise field. The model has the potential to popularize complex AI functions using consumer-grade hardware, which were once the exclusive research objects of technology giants and well-funded research institutions. This may help popularize AI-driven applications in various industries, such as optimizing customer service and providing more sophisticated data analysis tools.

2. OpenAI's efficient small model GPT-4o mini is affordable

OpenAI also launched GPT-4o mini, joining the competition for small models, which is known as the most cost-effective small model on the market. With only 15 cents per million tokens input and 60 cents per million tokens output, Venture Beat said that GPT-4o mini has greatly reduced the entry level of AI integration funds.

In addition, the pricing strategy of GPT-4o mini may spawn a new wave of AI-driven innovation, especially among startups and small businesses. By significantly reducing the cost of AI integration, the model effectively lowers the entry barrier to adopting AI-driven solutions. Ventrue Beat believes that this may accelerate technological innovation and reform in multiple industries. In addition, this shift to small models reflects a new trend in the AI ​​community: researchers and developers are increasingly concerned about efficiency, accessibility, and segmented applications. This trend may give rise to more targeted and efficient AI solutions that optimize specific tasks and industries.

3. Small models promote green technological innovation and reduce the carbon footprint of technology

The trend toward smaller models is also consistent with growing concerns about the environmental impact of AI. Small models consume less energy to train and run, potentially reducing the carbon footprint of AI technology. As technology companies face increasing pressure to be sustainable, Venture Beat believes that the low energy consumption and low emissions of small models may become an important selling point.

The environmental impact of this shift toward smaller models could be profound. As AI becomes more ubiquitous, the energy savings from more efficient models could be huge. This could allow AI to take a leading role in green innovation, rather than continuing to exacerbate global warming.

However, the rise of small models is not without challenges. As AI becomes more ubiquitous, questions of bias, accountability, and ethical use become more pressing. If not regulated, the popularization of AI through small models may amplify existing biases or create new ethical dilemmas. For developers and users of small models, ethical issues must be prioritized in addition to technical capabilities.

Conclusion: The future of AI development points to diversification and specialization

Although small models have advantages in efficiency and popularity, due to the limitation of parameter quantity, their processing capabilities on many tasks may not be comparable to large language models. Venture Beat said that this shows that in the future AI development landscape, there will be language models of various sizes, and smaller models will have their own specific areas of expertise.

Looking ahead, we expect to see a wide range of AI models, whether large language models or small models. A one-size-fits-all approach is not advisable. The key is to find a balance between model size, performance, and specific application requirements. For businesses and technology decision makers, these three small models represent a shift toward more efficient, professional, and deployable AI solutions, providing new possibilities for the integration of AI in enterprise environments. As for whether small models can replace the current dominance of large models in the cloud, it is too early to draw a conclusion.

Source: VentureBeat