GenAI

How to choose the right LLM for enterprise AI programs

Jul 15, 2024

5 min read

At the last count, there were over 2,350+ large language models in the market. Choosing one in two thousand can be daunting for any CIO. It doesn’t have to be, if you know what you’re looking for.

In this blog post, we give you the map for your enterprise AI treasure hunt. But first, some basics.

Each model is defined by a number of characteristics, like the LLM architecture, data it’s trained on, the output it produces, its cost, vendor, and the like. Before choosing the right LLM, it’s important to understand these characteristics.

Model development approach

LLMs are primarily of two types—closed-source and open-source. 

Closed-source is proprietary and maintained by the model vendor, who doesn’t make the source code publicly available. Popular models like ChatGPT, Google’s Gemini, and Anthropic’s Claude are closed-source. These models are advanced, rapidly evolving, up-to-date and easy to adapt to organizations of all kinds.

Open source democratizes development and access, allowing you to build on the model. Facebook’s Llama, Mistral, and Falcon are open-source. These models offer greater flexibility and cost-effectiveness.

Output type

The most common enterprise LLMs of today can generate text (like ChatGPT), images (like DALL-E), audio (like MusicLM), video (like Sora), and code (OpenAI Codex). Models that generate synthetic data, 3D models, etc., are also emerging.

Input mode

LLMs can be unimodal or multimodal. This means that users can prompt using just text or a combination of text, images, and videos. GPT-4 is multimodal.

Size

Don’t be fooled by the word ‘large’ in large language models. The size of the model depends on the number of parameters it is trained on. A model with a few billion parameters is considered small. Those with hundreds of billions of parameters, such as Google AI’s PaLM, are considered larger models.

Purpose

Some of the most popular enterprise LLMs are general-purpose, in that they cover a wide spectrum of knowledge and are trained on data available on the Internet. There are also domain-specific models trained in specialized languages in fields such as healthcare, finance, software engineering, etc.

With that in mind, let’s look at how you can choose the right model for your needs.

How do you choose the right LLM for enterprise AI use cases?

Making the right choice requires strategic thinking and AI expertise. Here’s a step-by-step approach.

1. Identify your use case

The model you use depends almost entirely on the purpose it needs to serve. So, select the use case and the best model for it. For example, if you’re automating workflows, you need a function-calling model. If you want to do data analysis, you need a search and retrieval model. 

2. Crystallize your business case

Use case is one thing - it defines what your LLM does. What’s more important for sustained success is a powerful business case. 

  • Focus on the ROI, set up systems to identify and measure key metrics

  • Calculate the cost per inference at which you make positive ROI

  • Consider leveraging multiple LLMs for a single task

  • Look for models that meet your ROI threshold

If you’re evaluating the expenses of implementing GenAI, try our TCO calculation guide here.

3. Focus on suitability

Not all use cases need large language models. Less is more. Sometimes, a small language model (SLM) that is fine-tuned or grounded in your data will outperform an LLM. 

For instance, DeepSeek Coder is a 3 billion parameter model - much smaller than GPT-4 - but exceptional at certain tasks. It even outperforms the latter in some instances.

4. Decide on open vs. closed models

While open-source models are cheaper in the long run, they might end up being much costlier to deploy and manage on your own. On the other hand, giving away your proprietary data to fine-tune a closed-source model for an enterprise GPT might cost you your competitive advantage!

Ask yourself the following questions:

  • Would you like to customize?

  • Is your data sensitive?

  • Is your data proprietary to you?

  • Do you have the bandwidth for fine-tuning, deploying and managing your models?

Based on your answers, choose your LLM and the correct license model for your GenAI roadmap.

5. Explore customizability

While most GenAI models work out of the box, they work best when customized using your data. This creates your own personalized enterprise GPT.

Imagine your marketing team writing a single prompt to curate a database of all customers who have made purchases, within the last three months, using a particular coupon code, in your e-commerce channel. Traditional CRMs would demand multiple filters and sorting. Your custom GenAI needs just a question!

Consider model customization. Choose the right data for the models to learn from. Invest in data quality initiatives throughout your LLM lifecycle. Experiment and test repeatedly to find the right recipe.

6. Prioritize accuracy

Your GenAI applications are only as useful as they are accurate. Not only should answers be accurate, they should be so on the first attempt. Here, concepts of zero-shot and multi-shot learning apply.

Zero-shot learning accepts prompts and produces output without needing examples or contextual training. This is less accurate for niche or domain-specific input. Multi-shot learning uses labeled examples to generate content resembling them closely. The output here is accurate and contextual.

Choose the method that best suits your needs. Then, conduct thorough testing with the right questions and evaluation for these models. 

7. Be strategic

The best part about GenAI implementation is that you can be agile, iterative, and build incrementally. So, be strategic in your Generative AI adoption.

Prototype first: If you’re just starting your GenAI journey, open source can be too much to handle. So, start prototyping on closed models and eventually switch to open source. 

Start small: Choose simple use cases before jumping into complex ones. Focus on delivering ROI and building confidence internally before expanding.

Engage people: Identify stakeholders and champions within the organization. Invite beta users to try the GenAI implementation and offer meaningful feedback. 

When in doubt, find a great partner. Tune AI’s experts have wide-ranging experience in custom models, open strack, infra control, data protection, security, and more. 

Speak to us for the best advice on choosing the right model for your enterprise AI solutions.

At the last count, there were over 2,350+ large language models in the market. Choosing one in two thousand can be daunting for any CIO. It doesn’t have to be, if you know what you’re looking for.

In this blog post, we give you the map for your enterprise AI treasure hunt. But first, some basics.

Each model is defined by a number of characteristics, like the LLM architecture, data it’s trained on, the output it produces, its cost, vendor, and the like. Before choosing the right LLM, it’s important to understand these characteristics.

Model development approach

LLMs are primarily of two types—closed-source and open-source. 

Closed-source is proprietary and maintained by the model vendor, who doesn’t make the source code publicly available. Popular models like ChatGPT, Google’s Gemini, and Anthropic’s Claude are closed-source. These models are advanced, rapidly evolving, up-to-date and easy to adapt to organizations of all kinds.

Open source democratizes development and access, allowing you to build on the model. Facebook’s Llama, Mistral, and Falcon are open-source. These models offer greater flexibility and cost-effectiveness.

Output type

The most common enterprise LLMs of today can generate text (like ChatGPT), images (like DALL-E), audio (like MusicLM), video (like Sora), and code (OpenAI Codex). Models that generate synthetic data, 3D models, etc., are also emerging.

Input mode

LLMs can be unimodal or multimodal. This means that users can prompt using just text or a combination of text, images, and videos. GPT-4 is multimodal.

Size

Don’t be fooled by the word ‘large’ in large language models. The size of the model depends on the number of parameters it is trained on. A model with a few billion parameters is considered small. Those with hundreds of billions of parameters, such as Google AI’s PaLM, are considered larger models.

Purpose

Some of the most popular enterprise LLMs are general-purpose, in that they cover a wide spectrum of knowledge and are trained on data available on the Internet. There are also domain-specific models trained in specialized languages in fields such as healthcare, finance, software engineering, etc.

With that in mind, let’s look at how you can choose the right model for your needs.

How do you choose the right LLM for enterprise AI use cases?

Making the right choice requires strategic thinking and AI expertise. Here’s a step-by-step approach.

1. Identify your use case

The model you use depends almost entirely on the purpose it needs to serve. So, select the use case and the best model for it. For example, if you’re automating workflows, you need a function-calling model. If you want to do data analysis, you need a search and retrieval model. 

2. Crystallize your business case

Use case is one thing - it defines what your LLM does. What’s more important for sustained success is a powerful business case. 

  • Focus on the ROI, set up systems to identify and measure key metrics

  • Calculate the cost per inference at which you make positive ROI

  • Consider leveraging multiple LLMs for a single task

  • Look for models that meet your ROI threshold

If you’re evaluating the expenses of implementing GenAI, try our TCO calculation guide here.

3. Focus on suitability

Not all use cases need large language models. Less is more. Sometimes, a small language model (SLM) that is fine-tuned or grounded in your data will outperform an LLM. 

For instance, DeepSeek Coder is a 3 billion parameter model - much smaller than GPT-4 - but exceptional at certain tasks. It even outperforms the latter in some instances.

4. Decide on open vs. closed models

While open-source models are cheaper in the long run, they might end up being much costlier to deploy and manage on your own. On the other hand, giving away your proprietary data to fine-tune a closed-source model for an enterprise GPT might cost you your competitive advantage!

Ask yourself the following questions:

  • Would you like to customize?

  • Is your data sensitive?

  • Is your data proprietary to you?

  • Do you have the bandwidth for fine-tuning, deploying and managing your models?

Based on your answers, choose your LLM and the correct license model for your GenAI roadmap.

5. Explore customizability

While most GenAI models work out of the box, they work best when customized using your data. This creates your own personalized enterprise GPT.

Imagine your marketing team writing a single prompt to curate a database of all customers who have made purchases, within the last three months, using a particular coupon code, in your e-commerce channel. Traditional CRMs would demand multiple filters and sorting. Your custom GenAI needs just a question!

Consider model customization. Choose the right data for the models to learn from. Invest in data quality initiatives throughout your LLM lifecycle. Experiment and test repeatedly to find the right recipe.

6. Prioritize accuracy

Your GenAI applications are only as useful as they are accurate. Not only should answers be accurate, they should be so on the first attempt. Here, concepts of zero-shot and multi-shot learning apply.

Zero-shot learning accepts prompts and produces output without needing examples or contextual training. This is less accurate for niche or domain-specific input. Multi-shot learning uses labeled examples to generate content resembling them closely. The output here is accurate and contextual.

Choose the method that best suits your needs. Then, conduct thorough testing with the right questions and evaluation for these models. 

7. Be strategic

The best part about GenAI implementation is that you can be agile, iterative, and build incrementally. So, be strategic in your Generative AI adoption.

Prototype first: If you’re just starting your GenAI journey, open source can be too much to handle. So, start prototyping on closed models and eventually switch to open source. 

Start small: Choose simple use cases before jumping into complex ones. Focus on delivering ROI and building confidence internally before expanding.

Engage people: Identify stakeholders and champions within the organization. Invite beta users to try the GenAI implementation and offer meaningful feedback. 

When in doubt, find a great partner. Tune AI’s experts have wide-ranging experience in custom models, open strack, infra control, data protection, security, and more. 

Speak to us for the best advice on choosing the right model for your enterprise AI solutions.

At the last count, there were over 2,350+ large language models in the market. Choosing one in two thousand can be daunting for any CIO. It doesn’t have to be, if you know what you’re looking for.

In this blog post, we give you the map for your enterprise AI treasure hunt. But first, some basics.

Each model is defined by a number of characteristics, like the LLM architecture, data it’s trained on, the output it produces, its cost, vendor, and the like. Before choosing the right LLM, it’s important to understand these characteristics.

Model development approach

LLMs are primarily of two types—closed-source and open-source. 

Closed-source is proprietary and maintained by the model vendor, who doesn’t make the source code publicly available. Popular models like ChatGPT, Google’s Gemini, and Anthropic’s Claude are closed-source. These models are advanced, rapidly evolving, up-to-date and easy to adapt to organizations of all kinds.

Open source democratizes development and access, allowing you to build on the model. Facebook’s Llama, Mistral, and Falcon are open-source. These models offer greater flexibility and cost-effectiveness.

Output type

The most common enterprise LLMs of today can generate text (like ChatGPT), images (like DALL-E), audio (like MusicLM), video (like Sora), and code (OpenAI Codex). Models that generate synthetic data, 3D models, etc., are also emerging.

Input mode

LLMs can be unimodal or multimodal. This means that users can prompt using just text or a combination of text, images, and videos. GPT-4 is multimodal.

Size

Don’t be fooled by the word ‘large’ in large language models. The size of the model depends on the number of parameters it is trained on. A model with a few billion parameters is considered small. Those with hundreds of billions of parameters, such as Google AI’s PaLM, are considered larger models.

Purpose

Some of the most popular enterprise LLMs are general-purpose, in that they cover a wide spectrum of knowledge and are trained on data available on the Internet. There are also domain-specific models trained in specialized languages in fields such as healthcare, finance, software engineering, etc.

With that in mind, let’s look at how you can choose the right model for your needs.

How do you choose the right LLM for enterprise AI use cases?

Making the right choice requires strategic thinking and AI expertise. Here’s a step-by-step approach.

1. Identify your use case

The model you use depends almost entirely on the purpose it needs to serve. So, select the use case and the best model for it. For example, if you’re automating workflows, you need a function-calling model. If you want to do data analysis, you need a search and retrieval model. 

2. Crystallize your business case

Use case is one thing - it defines what your LLM does. What’s more important for sustained success is a powerful business case. 

  • Focus on the ROI, set up systems to identify and measure key metrics

  • Calculate the cost per inference at which you make positive ROI

  • Consider leveraging multiple LLMs for a single task

  • Look for models that meet your ROI threshold

If you’re evaluating the expenses of implementing GenAI, try our TCO calculation guide here.

3. Focus on suitability

Not all use cases need large language models. Less is more. Sometimes, a small language model (SLM) that is fine-tuned or grounded in your data will outperform an LLM. 

For instance, DeepSeek Coder is a 3 billion parameter model - much smaller than GPT-4 - but exceptional at certain tasks. It even outperforms the latter in some instances.

4. Decide on open vs. closed models

While open-source models are cheaper in the long run, they might end up being much costlier to deploy and manage on your own. On the other hand, giving away your proprietary data to fine-tune a closed-source model for an enterprise GPT might cost you your competitive advantage!

Ask yourself the following questions:

  • Would you like to customize?

  • Is your data sensitive?

  • Is your data proprietary to you?

  • Do you have the bandwidth for fine-tuning, deploying and managing your models?

Based on your answers, choose your LLM and the correct license model for your GenAI roadmap.

5. Explore customizability

While most GenAI models work out of the box, they work best when customized using your data. This creates your own personalized enterprise GPT.

Imagine your marketing team writing a single prompt to curate a database of all customers who have made purchases, within the last three months, using a particular coupon code, in your e-commerce channel. Traditional CRMs would demand multiple filters and sorting. Your custom GenAI needs just a question!

Consider model customization. Choose the right data for the models to learn from. Invest in data quality initiatives throughout your LLM lifecycle. Experiment and test repeatedly to find the right recipe.

6. Prioritize accuracy

Your GenAI applications are only as useful as they are accurate. Not only should answers be accurate, they should be so on the first attempt. Here, concepts of zero-shot and multi-shot learning apply.

Zero-shot learning accepts prompts and produces output without needing examples or contextual training. This is less accurate for niche or domain-specific input. Multi-shot learning uses labeled examples to generate content resembling them closely. The output here is accurate and contextual.

Choose the method that best suits your needs. Then, conduct thorough testing with the right questions and evaluation for these models. 

7. Be strategic

The best part about GenAI implementation is that you can be agile, iterative, and build incrementally. So, be strategic in your Generative AI adoption.

Prototype first: If you’re just starting your GenAI journey, open source can be too much to handle. So, start prototyping on closed models and eventually switch to open source. 

Start small: Choose simple use cases before jumping into complex ones. Focus on delivering ROI and building confidence internally before expanding.

Engage people: Identify stakeholders and champions within the organization. Invite beta users to try the GenAI implementation and offer meaningful feedback. 

When in doubt, find a great partner. Tune AI’s experts have wide-ranging experience in custom models, open strack, infra control, data protection, security, and more. 

Speak to us for the best advice on choosing the right model for your enterprise AI solutions.

Written by

Al Rey

GTM Advisor