Want to make your own AI Agent? Collect these 12 LLM models, and you can also train good tools!

I receive similar questions almost every day. After helping to build more than 20 AI intelligences and investing a lot of costs in testing models, I have summarized some truly effective experiences.

Here is the complete guide on how to choose the right LLM.

Image Source: TechFlow Shenzhen

The current Large Language Model (LLM) industry is changing rapidly. There are new models released almost every week, and each model claims to be the 'best'.

But the reality is: there is no model that can meet all needs.

Each model has its specific applicable scenario.

I have tested dozens of models and hope that through my experience, I can help you avoid unnecessary waste of time and money.

Image Source: TechFlow Shenzhen

It should be noted that this article is not based on laboratory benchmark tests or marketing promotions.

I will share the practical experience of building AI intelligent body and generative AI (GenAI) products over the past two years.

First, we need to understand what LLM is:

The Large Language Model (LLM) is like teaching a computer to "speak human." It predicts the most likely words to appear next based on the content you input.

The starting point of this technology is this classic paper: Attention Is All You Need.

Fundamental Knowledge - LLM of Closed Source Code and Open Source Code:

Closed-source code: For example, GPT-4 and Claude, are often paid for based on usage and hosted by the provider to run.

Open source: for example, Meta's Llama and Mixtral need to be deployed and run by users themselves.

When you first encounter it, you may be confused by these terms, but it is important to understand the difference between the two.

Image Source: TechFlow Shenzhen

Model size does not necessarily mean better performance:

For example, 7B indicates that the model has 7 billion parameters.

But a larger model doesn't always perform better. The key is to choose a model that suits your specific needs.

Image Source: TechFlow Shenzhen

If you need to build X/Twitter bots or social AI:

@xai's Grok is a very good choice:

Provide generous free quota

Outstanding ability to understand social context

Although it is closed source code, it is worth trying.

Strongly recommend this model for novice developers! (Rumor:

@ai16zdao's Eliza default model is using XAI Grok）

If you need to handle multilingual content:

@Alibaba_Qwen's QwQ model performed very well in our tests, especially in Asian language processing.

It should be noted that the training data of this model mainly comes from mainland China, so there may be information missing in certain content.

Image Source: TechFlow Shenzhen

If you need a general-purpose or strong reasoning model:

@OpenAI 的模型依然是業界的佼佼者：

Stable and reliable performance

Extensively tested in practice

Has a strong security mechanism

This is the ideal starting point for most projects.

Image Source: TechFlow Shenzhen

If you're a developer or content creator:

Claude from @AnthropicAI is the main tool I use on a daily basis:

The coding ability is quite outstanding

The response is clear and detailed

Great for creative-related work

Image Source: TechFlow Shenzhen

Meta's Llama 3.3 has recently attracted attention:

Stable and reliable performance

Open source model, flexible and free

You can try it out through @OpenRouterAI or @GroqInc

For example, @virtuals_io and other cryptocurrency x AI projects are developing products based on it.

Image Source: TechFlow Shenzhen

If you need role-playing AI:

@TheBlokeAI's MythoMax 13B is currently a top player in the role-playing industry, consistently ranking high in relevant rankings for several months.

Cohere's Command R+ is an underrated excellent model:

Performed well in role-playing missions

Able to easily handle complex tasks

Supports up to 128,000 contextual windows with longer "memory capabilities"

Image Source: TechFlow Shenzhen

Google's Gemma model is a lightweight but powerful choice:

Focused on specific tasks, performs excellently

Budget-friendly

Suitable for cost-sensitive projects

Personal experience: I often use the small Gemma model as an 'unbiased judge' in the AI process, and it works very well in verification tasks!

Image Source: TechFlow Shenzhen

Gemma

@MistralAI 的模型值得一提：

Open source with high-end quality

The performance of Mixtral models is very powerful.

He is particularly good at complex reasoning tasks

It has received widespread praise from the community and is definitely worth a try.

Cutting-edge AI in your hands.

Pro tip: Try mixing and matching!

Different models have their own advantages

You can create AI 'teams' for complex tasks

Let each model focus on its strengths

It's like building a dream team, each member has a unique role and contribution.

How to get started quickly:

Use @OpenRouterAI or @redpill_gpt for model testing. These platforms support cryptocurrency payment, which is very convenient.

It is an excellent tool for comparing the performance of different models.

If you're looking to save costs and run your model locally, you can try experimenting with your own GPU using @ollama.

Image Source: TechFlow Shenzhen

If you're looking for speed, @GroqInc's LPU technology offers extremely fast inference speeds:

Although the model selection is limited

However, performance is well suited for deployment in a production environment

Image Source: TechFlow Shenzhen

[Disclaimer] The market is risky, and investment needs to be cautious. This article does not constitute investment advice, and users should consider whether any opinions, views, or conclusions in this article are in line with their specific situation. Invest at your own risk.

This article is reprinted with permission from: Deep Tide TechFlow

Original Author: superoo7

『Want to make your own AI Agent?』 12 LLM models to collect, you can also tune good tools!' This article was first published in 'Crypto City'