June Top Models
Welcome to our first installments of monthly Top Models. We aim to make choosing models to work with as easy as possible.
We will start with two leaderboards; top models and open source. Top Models are our strongest recommendation when building Generative AI solutions for personal and commercial applications.
Our Open Source leaderboard provides insights into our recommendations for best models to use as a starting point when working on customized generative models. They might not outperform our top models across a multitude of benchmarks, but they can be fine tuned into performing strongly and cheaply on tasks you care most about.
OpenAI GPT4o: Well balanced price for performance. OpenAI takes the top spot with this multimodal, multilingual, impressive offering. We find 4o performance across just about any task to be strong.
OpenAI GPT4: Like father like son, GPT4 takes second place for all the same reasons as 4o. When it comes to more complicated tasks we recommend this model however for most uses this is needlessly expensive.
Anthropic Claude Opus: Challenger Approaching! It may not dethrone OpenAI however when it comes to generated text that doesn’t sound synthetic and writing complex code, Claude is our top choice.
WizardLM 2: Microsoft has been on a tear with model releases and this one should be on everyone’s short list. Flavors of 7B, 70B, and 8x22B with an Apache 2 license makes this model accessible to all. We can quote the whitepaper specs but it’s best explored on your own; you’ll thank us.
Mistral v3: The new Mistral hotness. While it only comes in a 7B flavor, it features an Apache 2 license and function calling.
Phi3: Who says a 3B parameter model can’t punch up. Microsoft has been actively demonstrating the power of small but mighty models that can be cost effectively trained and deployed to your needs.
Llama3: Meta never disappoints with the performance of their Llama model releases. Llama 3 is no exception with their 8B and 70B flavors. The challenge? It’s license requires you to indemnify Meta for any uses of their model. That makes is great for personal use but probably a deal breaker for any corporation.
Mixtral v3 8x22B: Cue the John Cena meme. It doesn’t even have a model card yet but we see you @ https://huggingface.co/mistral-community/mixtral-8x22B-Instruct-v0.3-original
Cultivation
It all begins with an idea.
AI has vaulted from academia to the boardroom but it still hasn’t landed gracefully yet. No matter your level of expertise, consuming information on the domain is overwhelming; further exacerbated by every product incorporating some form of generative AI or MLOps strategy into their offering. Distilling bs from brilliance can be difficult.
Don’t worry, we’ve got you. You don’t need to scour thousands of models on huggingface, read thousands of papers on arxiv, or sift through a myriad of product pitches to find out it’s just wrapping GPT.
We are going, in parallel, to follow two paths of research:
Curating information for the curious
Cultivating information for the informed
For those of you active in the field, our goal is to simplify the white paper noise and model release hype. Through both social media and our internal library we will source the best information to stay apprised of.
But don’t worry if you aren’t embedding all the things yet. It doesn’t matter if you are a software engineer looking to learn prompting, an executive hoping to make informed choices on commercial models, or just excited for the technology; we have plans to consolidate your journey to meaningful information.