The modern LLM landscape

Artificial intelligence (AI) has long been a subject of fascination and research, but recent advancements in generative AI have propelled it into mainstream adoption. Unlike traditional AI systems that classify data or make predictions, generative AI can create new content—text, images, code, and more—by leveraging vast amounts of training data.

The generative AI revolution was catalyzed by the 2017 introduction of the transformer architecture, which enabled models to process text with unprecedented understanding of context and relationships. As researchers scaled these models from millions to billions of parameters, they discovered something remarkable: larger models didn’t just perform incrementally better—they exhibited entirely new emergent capabilities like few-shot learning, complex reasoning, and creative generation that weren’t explicitly programmed. Eventually, the release of ChatGPT in 2022 marked a turning point, demonstrating these capabilities to the public and sparking widespread adoption.

The landscape shifted again with the open-source revolution led by models like Llama and Mistral, democratizing access to powerful AI beyond the major tech companies. However, these advanced capabilities came with significant limitations—models couldn’t reliably use tools, reason through complex problems, or maintain context across interactions. This gap between raw model power and practical utility created the need for specialized frameworks like LangChain that transform these models from impressive text generators into functional, production-ready agents capable of solving real-world problems.

Table 1.1: A timeline of major developments in language models

The field of LLMs is rapidly evolving, with multiple models competing in terms of performance, capabilities, and accessibility. Each provider brings distinct advantages, from OpenAI’s advanced general-purpose AI to Mistral’s open-weight, high-efficiency models. Understanding the differences between these models helps practitioners make informed decisions when integrating LLMs into their applications.

Model comparison

The following points outline key factors to consider when comparing different LLMs, focusing on their accessibility, size, capabilities, and specialization:

• Open-source vs. closed-source models: Open-source models like Mistral and LLaMA provide transparency and the ability to run locally, while closed-source models like GPT-4 and Claude are accessible through APIs. Open-source LLMs can be downloaded and modified, enabling developers and researchers to investigate and build upon their architectures, though specific usage terms may apply.
• Size and capabilities: Larger models generally offer better performance but require more computational resources. This makes smaller models great for use on devices with limited computing power or memory, and can be significantly cheaper to use. Small language models (SLMs) have a relatively small number of parameters, typically using millions to a few billion parameters, as opposed to LLMs, which can have hundreds of billions or even trillions of parameters.
• Specialized models: Some LLMs are optimized for specific tasks, such as code generation (for example, Codex) or mathematical reasoning (e.g., Minerva).

The increase in the scale of language models has been a major driving force behind their impressive performance gains. However, recently there has been a shift in architecture and training methods that has led to better parameter efficiency in terms of performance.

Let’s now discuss a comparative overview of various LLMs, highlighting their key characteristics and differentiating factors. We’ll delve into aspects such as open-source vs. closed-source models, model size and capabilities, and specialized models. By understanding these distinctions, you can select the most suitable LLM for your specific needs and applications.

LLM provider landscape

You can access LLMs from major providers like OpenAI, Google, and Anthropic, along with a growing number of others, through their websites or APIs. As the demand for LLMs grows, numerous providers have entered the space, each offering models with unique capabilities and trade-offs. Developers need to understand the various access options available for integrating these powerful models into their applications. The choice of provider will significantly impact development experience, performance characteristics, and operational costs.

The table below provides a comparative overview of leading LLM providers and examples of the models they offer:

Table 1.2: Comparative overview of major LLM providers and their flagship models for LangChain implementation

Other organizations develop LLMs but do not necessarily provide them through application programming interfaces (APIs) to developers. For example, Meta AI develops the very influential Llama model series, which has strong reasoning, code-generation capabilities, and is released under an open-source license.

There is a whole zoo of open-source models that you can access through Hugging Face or through other providers. You can even download these open-source models, fine-tune them, or fully train them. We’ll try this out practically starting in Chapter 2.

Once you’ve selected an appropriate model, the next crucial step is understanding how to control its behavior to suit your specific application needs. While accessing a model gives you computational capability, it’s the choice of generation parameters that transforms raw model power into tailored output for different use cases within your applications.

Now that we’ve covered the LLM provider landscape, let’s discuss another critical aspect of LLM implementation: licensing considerations. The licensing terms of different models significantly impact how you can use them in your applications.

Licensing

LLMs are available under different licensing models that impact how they can be used in practice. Open-source models like Mixtral and BERT can be freely used, modified, and integrated into applications. These models allow developers to run them locally, investigate their behavior, and build upon them for both research and commercial purposes.

In contrast, proprietary models like GPT-4 and Claude are accessible only through APIs, with their internal workings kept private. While this ensures consistent performance and regular updates, it means depending on external services and typically incurring usage costs.

Some models like Llama 2 take a middle ground, offering permissive licenses for both research and commercial use while maintaining certain usage conditions. For detailed information about specific model licenses and their implications, refer to the documentation of each model or consult the model openness framework: https://isitopen.ai/.

In general, open-source licenses promote wide adoption, collaboration, and innovation around the models, benefiting both research and commercial development. Proprietary licenses typically give companies exclusive control but may limit academic research progress. Non-commercial licenses often restrict commercial use while enabling research.

By making knowledge and knowledge work more accessible and adaptable, generative AI models have the potential to level the playing field and create new opportunities for people from all walks of life.

The evolution of AI has brought us to a pivotal moment where AI systems can not only process information but also take autonomous action. The next section explores the transformation from basic language models to more complex, and finally, fully agentic applications.