Top AI Models

The Best Open Source and Closed Source (Proprietary) AI Models

Aug 17, 2024
 

Read time: 7 minutes

 


 

Struggling to keep up with all the AI models and updates? 

And, figure out what’s what.

Don’t worry, we feel you.

Right here, you’ll find the definitive list of the top AI models on the market today, both open-source and proprietary. And trust us, there are a few surprises in store. 

Whether you’re deep in the AI game or just starting to explore, this list will not only introduce you to the best models available but might just challenge what you think you know.

By the end of this article you will understand:

  1. The top models today
  2. How to compare models
  3. Which one's are most applicable for businesses

 

First - What Is Open Source?

Lets just start by attempting to define what is open source or proprietary in model terms. These categories reflect different philosophies and approaches to accessibility, transparency, and control over AI models.

 

1. Open Source Models

 

Definition:
Open-source models are those where the model’s source code, and often its architecture and weights, are made publicly available. This transparency allows anyone to inspect, modify, and redistribute the model. Open-source models can be further divided into two subcategories:

 

Subcategories:

  • Fully Open Source:
    These models provide full access to all aspects of the model, including the source code, architecture, weights, and sometimes even the training data. Users have the ability to fully replicate, modify, and build upon the model. This level of openness fosters innovation and collaboration within the AI community.

  • Partially Open Source:
    Some models are released under an open-source licence but do not make all components available. For instance, the model’s code and weights might be accessible, but the training data or certain architectural details remain proprietary. This approach balances community involvement with the protection of intellectual property.

Key Features of Open Source Models:

 

  • Transparency: Users can see and understand how the model works, including its architecture and, in some cases, the data it was trained on.
  • Customisability: Full or partial access allows users to modify the model to fit their specific needs.
  • Community Collaboration: Open-source models often benefit from contributions by a global community of developers, leading to continuous improvement and innovation.

 

 

2. Proprietary Models

 

Definition:
Proprietary models are fully controlled by the organisations that develop them. Access to these models is typically provided through their own interfaces or APIs under strict licensing agreements. The source code, architecture, weights, and training data are not shared with the public, and users are limited to the functionalities provided by the company.

 

Key Features:

  • High Performance: Proprietary models are often optimised for specific tasks and are maintained by dedicated teams, ensuring high levels of performance and reliability.
  • Security and Support: These models come with professional support and robust security features, managed by the developing organisation.
  • Limited Transparency and Customisation: Users have no access to the internal workings of the model and limited ability to modify or adapt it beyond the parameters provided by the API. In some instances however you can fine tune models so there is scope for partial customisation.

 

The landscape is split across Open Source and Proprietary models, each varying with the levels of transparency, control, and access offered. 

Fully open-source models provide maximum flexibility and community-driven innovation, while partially open-source models offer a compromise between openness and protecting proprietary elements. Proprietary models, on the other hand, deliver optimised performance and strong support but at the cost of transparency and customisation.

OK, now let's look at the best models out there today and what they do.

 

Top Open Source Models

 

 

  1. Qwen2 (Alibaba Cloud)

What?! Alibaba is building AI? Yes, Alibaba, the Chinese equivalent of Amazon.com, is building its own models. And they’re good.

  • Model Size (parameters): Up to 72B
  • Context Window: 128K tokens
  • Strengths: Scores high in General Ability and Reasoning & Knowledge (MMLU). In addition to quality against price compared to other top models. Robust multilingual capabilities, proficient in approximately 30 languages..
  • Weaknesses: It’s not bad but some way off on the Coding (HumanEval) scores compared to the leaders (Claude 3.5 Sonnet and GTP-4o). It’s also some way off in terms of speed across both tokens per second and latency (time to first token). Resource-intensive, particularly for the larger models, requiring specialised hardware for optimal performance..
  • Applications: Advanced multilingual content generation, coding, secure AI outputs.
  • Overall Score: 9.5/10

 

  1. Grok-2 (xAI)

I am Grok…get it? Guardians of….ahh never mind.

  • Model Size (parameters): 175B
  • Context Window: 128K tokens
  • Strengths: Strong in multilingual tasks, code generation, and ethical AI alignment.
  • Weaknesses: Slightly smaller parameter size than some competitors.
  • Applications: Conversational AI, content creation, code generation.
  • Overall Score: 9.5/10

 

  1. Nemotron-4 340B (Nvidia)

My new favourite transformer alongside Optimus Prime.

  • Model Size (parameters): 340B
  • Context Window: 128K tokens
  • Strengths: Excels in high-precision tasks and AI safety.
  • Weaknesses: Very resource-intensive.
  • Applications: Synthetic data generation, AI safety, complex NLP tasks.
  • Overall Score: 9.5/10

 

  1. Yi-1.5 (Alibaba Cloud)

Nope it’s not the name of Elon’s next child, it’s another Alibaba model specialising in multilingual applications.

  • Model Size (parameters): 15B
  • Context Window: 128K tokens
  • Strengths: Strong in multilingual tasks and general-purpose NLP.
  • Weaknesses: Slightly lower in coding and mathematical reasoning.
  • Applications: Customer support, multilingual processing.
  • Overall Score: 9/10

 

  1. LLaMA 3 (Meta)

“It’s not the size of your context window it’s what you do with it, alright!” - LLaMA 3

  • Model Size (parameters): 8B, 70B, 400B (upcoming)
  • Context Window: 8,192 tokens
  • Strengths: Versatile in general-purpose NLP and multilingual support.
  • Weaknesses: Smaller context window; may need fine-tuning.
  • Applications: Conversational AI, research.
  • Overall Score: 9/10

 

  1. Mistral Large 2 (Mistral AI)

A chatbot’s a chatbot unless you’re Mistal and then it’s “Le Chat”

  • Model Size (parameters): 123B
  • Context Window: (Not explicitly defined)
  • Strengths: High precision in technical tasks.
  • Weaknesses: Resource-heavy, lacks long-context handling.
  • Applications: Technical document processing, research.
  • Overall Score: 8.5/10

 

Open Source Summary:

  • Top Performers: Qwen2, Grok-2, and Nemotron-4 excel across various domains.
  • General Purpose: Yi-1.5 and LLaMA 3 are strong in multilingual and general-purpose tasks.
  • Specialized: Mistral Large 2 is ideal for high-precision tasks but less versatile.

 

 

Top Proprietary Models

 

  1. GPT-4o (OpenAI)

GPT-4o: The golden protocol droid of AI, just as fluent in languages, but without the stiff joints and constant reminders of impossible odds.

  • Model Size (parameters): Confidential (100B+)
  • Context Window: 128K tokens
  • Strengths: Versatile across various NLP tasks.
  • Weaknesses: API-only access; no model weights.
  • Applications: Chatbots, content creation, research.
  • Overall Score: 9.5/10

 

  1. Gemini 1.5 Pro (Google)

Gemini “has everyone seen the size of my context window” 1.5 Pro

  • Model Size (parameters): Estimated 100B+
  • Context Window: 2 million tokens
  • Strengths: Optimised for enterprise-level tasks.
  • Weaknesses: Restricted to Google’s ecosystem.
  • Applications: Enterprise AI, complex data analysis.
  • Overall Score: 9.5/10

 

  1. Claude 3.5 Sonnet (Anthropic)

Ethics? Check. Conversations? Nailed it. Who needs the biggest context window when you've got the biggest conscience?

  • Model Size (parameters): Confidential
  • Context Window: 128K tokens
  • Strengths: Ethical AI, strong alignment tasks.
  • Weaknesses: Slightly less versatile in general tasks.
  • Applications: Ethical AI, conversational AI.
  • Overall Score: 9/10

 

  1. Cohere Command R (Cohere AI)

The “University Professor”, focused on research and non-commercial uses 

  • Model Size (parameters): 35B, 104B (Command R+)
  • Context Window: 128K tokens
  • Strengths: Specialised in RAG and multilingual tasks.
  • Weaknesses: Focused on specific tasks.
  • Applications: Search, knowledge management.
  • Overall Score: 9/10

 

  1. Jurassic-2 (AI21 Labs)

Bringing life to your ideas, no DNA extraction required. 

Note: AI21 Labs have released a new model called Jamba which is open source and outperforms this one.

  • Model Size (parameters): 178B
  • Context Window: 128K tokens
  • Strengths: Creative problem-solving, content generation.
  • Weaknesses: Resource-intensive.
  • Applications: Creative content generation, large-scale NLP.
  • Overall Score: 9/10

 

  1. PaLM 2 (Google)

PaLM 2: Your friendly neighbourhood Google-powered AI, with smarts.

  • Model Size (parameters): 340B
  • Context Window: 128K tokens
  • Strengths: Capable in reasoning, coding, and multilingual tasks.
  • Weaknesses: Limited to Google's platforms.
  • Applications: Reasoning, coding, multilingual 
  • Overall Score: 9/10

 

Proprietary Summary:

  • Top Performers: GPT-4o, Gemini 1.5 Pro, and Claude 3.5 Sonnet are the leaders in their respective domains.
  • Specialised: Cohere Command R and Jurassic-2 excel in niche applications.
  • Versatile: PaLM 2 offers robust capabilities within Google's ecosystem.

 

 

Want More Including Usage Costs and Speed?

If you want a more detailed breakdown of these models that includes information around their licence agreements, usage costs, quality and speed then get in touch here and we’ll send it to you. 

 

Conclusion

Navigating the complex landscape of AI models can feel overwhelming, but understanding the key differences between open source and proprietary models is essential for making the right choice for your business. 

Open-source models offer transparency and flexibility, empowering you to modify and innovate freely. If you run them within your own infrastructure they also give you the option of using them to process sensitive data.

On the other hand, proprietary models deliver optimised performance and more support, as well as potentially integrating with existing systems more easily.

Both types have their strengths, and the best choice depends on your specific needs. Are you looking for a model that you can tailor and adapt, or do you need a powerful, ready-to-use solution backed by professional support? Understanding these trade-offs will help you align the right AI model with your strategic goals.

 

Ready to harness the power of AI for your business?

Whether you're considering the flexibility of open-source models or the performance of proprietary solutions, making an informed choice is crucial. 

If you need further guidance or want to discuss how these models can be integrated into your strategy, don't hesitate to get in touch with our team who will be able to discuss what’s the right AI solution for your business.

Stay Ahead with The Ultimate AI Newsletter

Subscribe forĀ unique AI insightsĀ andĀ strategiesĀ that redefine business and innovation. Plus, get VIP access to a curated selection of "bad AI" - because sometimes, learning what not to do is just as valuable..