You know the gold standard by now.  Initially released over a year ago GPT still stands at the top of many benchmark scores.  With turbo, vision, and fine tuning options available (between Microsoft and OpenAI), GPT4 is an easy place to start for all your generative AI needs.

Anthropic's Claude series of models features three different model sizes available for access on Anthropics website or through Amazon's AWS Bedrock services.  Claude is cost effective, fast, and is our recommendation for a base model to try for RAG implementations.

The founders of Transformers may not always be making big news with their model releases but Gemini on GCP isn't a bad option.  With Gemini 1.5 you can pass in 1M tokens in an input, which may eliminate the need for a RAG solution depending on your use-case.

Founded by former Google DeepMind and Meta employees, Mistral has the pedigree and funding to challenge leaders in the AI space.  Their largest model, Mistral Large, is available only on Azure and AWS platforms.  It might not beat GPT4, but everyone should be paying attention to them.

We havent had the chance to play around with Cohere yet, but their Command model performs strongly in benchmarks and the companies philosophy resonates.  They are looking to balance quality, cost, and speed along with creating easy to use Enterprise tools to improve LLM oriented solutions.

The new king of open-ish (not Apache or MIT) models, Meta has outperformed the competition in both its 8B and 70B model variants.  The 8B model outperforms Mistral and Gemma, while the 70B outperforms Gemini Pro and Mixstral.  If you are looking to train a model of substance, this is where to start today.

Microsoft had to pull their release of Wizard 2 due to an internal compliance issue but the model has multiple flavors offering impressive performance.  Will update this row once it is formally published again.

A massive 132B parameter model trained on 12T tokens, DBRX is Databrick's and MosaicML's most substantial model.  While it still doesnt overthrow GPT4 for most use cases it performs staggeringly strong against human evaluated programming examples.  Databricks' MosaicML acquisition is likely to continue to result in continued research we are excited to see.

Using the Mixture of Experts approach to model architecture, Mixstral is an 8x7B parameter model that smooshes most open source competition.  Of course, its size may make it unsuitable for many use cases, it is still a remarkable release.  Licensed under Apache 2, it's yours to do as you please!

Using the architecture of Llama 2, TinyLlama is a different entry on this list.  At only 1.1B parameters it aims to maximize the value of each model weight.  By training TinyLlama against 3T tokens it continues to improve at each checkpoint.  When playing around at home, this is the model to start with.

When stacked against more recent open source models, Google's 2B and 7B Gemma is no longer relevant for most use cases.  That being said, it may still have value for specific domains and is worth keeping on this list as Google's family of more open to use models.

Mistral's 7B parameter model is a candle to the sun that is Mixstral but a great base model to continue to train on.

When it initially launched Falcon led the open source leaderboards.  Featured models at 7, 40, and 180 billion parameters Falcon made a tremendous splash.  Now, it's a piece of history more than a recommendation to evaluate.

A refreshed version of Falcon trained on over 5 Trillion tokens from RefinedWeb, a high quality dataset created by the Falcon team.  Falcon is back on the leaderboards.

Open-Orca 7B is less about the model, and more about the approach.  Using Mistral 7B as a base and their own OpenOrca dataset generated using GPT4 as a foundation to generate high quality synthetic data.  This is the foundation to what we believe is the cost effective model training approach of the future.   

Baked into ChatGPT, DALL-E 3 is OpenAI's text-to-image diffusion model that is capable of supporting highly detailed prompts, can render text better than any alternative, and scores the highest on human benchmarks.  While DALL-E 3 doesnt have all the built in commercial tools Stability and Midjourney provide, it is the best model to use when looking for an image with specific criteria.

Stability AI ushed in a wave of incredible text to image diffusion models with their initial Stable Diffusion release.  Since it's release there have been hundreds of open source models created for specific domains.  Stability itself has closed access to it's most recent version 3 release, exposing that only via API.  Stable Diffusion 2 is available as an MIT license and the results are still striking.  Tools such as Automatic1111 make running a stable diffusion model on your local machine a breeze.

Raise your hand if you had 'run one of the best AI text to image generators using Discord as a platform".  Well, it's true, and Midjourney has some impressive model results.  Using Discord as a platform, you can prompt for images across a suite of models inclusive of their Niji anime-style model in partnership with Spellbrush.

Beating OpenAI's ADA embedding model and released under Apache, it doesnt hurt to check the performance of this 137m parameter model with 768 dimentions supporting over 8k tokens. 

The Beijing Academy of AI released a series of multi-lingual embedding models are perform strongly are are decently small.

OpenAI is not leading the pack when it comes to embedding models but it does have an offering it continues to refine.  ADA isn't bad, but its not our recommendation for optimizing your RAG architecture

Starcoder isn't topping the charts for code generation models but it's Apache license allows you to use it as a baseline for continued pre-training on your repositories.  We would generally recommend GPT4 or Claude3 as a coding base but pay attention here.

Captured on our Products page, Whisper is OpenAI's speech to text transcription model supporting multiple language and model sizes to fit your needs.  If you plan to use this in production we recommend looking at hosting providers that serve up Whisper at scale and for likely as cheap as you would self hosting.  We see lots of companies selling transcription and it's nothing more than glossy UIs over Whisper.

Gorilla isn't making headlines yet, but leveraging fit-for-purpose generative models is likely the future we will all be living in.  When composing AI agents together, LLM's like Gorilla will be the binding agent.  We will have a blog post about this in the future.

OpenAI's MIT Licensed general purpose speech recognition model is an absolute beast.  With multi-lingual support, strong performance, and multiple model sizes to choose from, Whisper is the best open source transcription model to start with.  Many companies host implementations of Whisper making it a simple API call.

Whisper

AssemblyAI offers their own model, Universal-1 trained on a staggering 12.5M hours of multilingual audio. It claims to have the highest accuracy in the industry and benchmarks back it up.  Additionally, Universal-1 API performance is extremelly fast, capable of transcribing an hour of audio in about a minute. 

AssemblyAI

Unstructured is a collection of capabilities, both API/SaaS and Open Source, which assist in processing documents.  Outputs from Unstructured may be useful in training AI models, serving up data for RAGs, or useful in extracting information for direct processing by non AI systems.  Supporting a wide range of documents and platforms connections it is worth a look to see how much time you can save by processing documents with their capabilities. 

Unstructured

Where data processing competitors focus on the data, LlamaIndex offers a suite of capabilities to take data from parsing, to ingestion, to retrieval by enabling capabilities to support applications based on LLMs.  LlamaIndex is thinking broadly about enabling their capabilities for Enterprise use to hopefully make these types of processes simple in the future. 

LlamaIndex

Initially a hardware company focused on building optimized processors ideal for LLMs, Groq has now pivoted as an inferencing provider.  We can spout off a bunch of metrics as to how much faster their LPU is at hosting models but the results speak for themselves.  Go try out their chat interface and enjoy how quickly results are returned.  We haven't used their LPUs yet, but color us impressed so far.

Groq

Fireworks provides serverless hosting of many of the top open source AI model.  They also allow customers to fine tune LoRAs for use on their platform to tailor serverless models to your specific needs.  For Enterprise customers they also offer dedicated deployments using A100s.    

Fireworks

A model hosting platform with numerous imagery, text, and multimodal modals.  Deepinfra also offers 1 billion free tokens as part of their DeepStart program for startups meeting certain criteria. Deploying your own models are possible offering staggering prices for access to A100/H100s.  We have used Deepinfra for generating synthetic data from license free open source models and loved the experience and price.

Deepinfra

In the early days of Google Colab you could log in and be granted access to an nVidia Tesla V100; no longer.  Web enabled access to compute (GPU/CPU) is so delightful when trying to load up a notebook for experiment or model training.  RunPod is our recommendation as the best of the best.  We have used it extensively and can report that we have had no issues getting the GPUs we need.

RunPod

One of the best free courses to take to learn fundamental. Slightly dated now but still worth a look

Github Repo of content to take the curious from the foundations of LLM to building applications

Huggingface tutorial on fine tuning an open source model. Easy place to start the journey

Open Source tool to assist in generating synthetic data

AI & ML Resources