The best language models for digital products

Based on real benchmark data from our own software products, we evaluated the performance of different LLM models in addressing specific challenges. We examined specific categories such as document processing, CRM integration, external integration, marketing support, and code generation.  

Highlights:

• Gemini Pro 1.5 from Google - Improvement of Pro 1.0, now available in the EU

• Command-R and Command-R Plus from Cohere - mediocre results

• New GPT-4 Turbo - OpenAI has done it again!

• Llama 3: 70B is fine, but 8B is really promising

• Long-term trends

Our benchmarks evaluate the models in terms of their suitability for digital product development. The higher the score, the better.

☁️ - Cloud models with proprietary license
✅ - Open source models that can be run locally without restrictions
🦙 - Local models with Llama2 license

A more detailed explanation of the respective categories can be found below the table.

Google Gemini Pro 1.5

The newer version, Gemini 1.5 Pro shows significantly better performance than Gemini 1.0 Pro in the last month. It almost reaches the performance of the GPT-4 Turbo.

This model performs particularly well on tasks related to working on documents and information. It also performs almost perfectly on CRM-related tasks. However, complex reasoning tasks are below the level of GPT-3.5.

Gemini Pro 1.5 is about 20 times more expensive than Pro 1.0 on our workloads, which is to be expected given the quality level of GPT-4.

Both models are now available in Google Vertex AI, finally making them usable for enterprise customers in the EU.

Command R models from Cohere

Cohere AI specializes in enterprise-oriented LLMs. They have the Command R family of models - LLMs designed for document-oriented tasks: "Command R" and "Command R Plus".

These models are available both as API SaaS and as downloadable models on Hugging Face. Downloadable models are published for non-commercial purposes.

The Command-R model is roughly comparable to the Anthropic Claude models of the first two generations, but significantly cheaper. Nevertheless, there are better models in this price category, such as Gemini Pro 1.0 and Claude 3 Haiku.

The Command R+ is a significantly better model with capabilities in the GPT-3.5 range, but at two to three times the price.

OpenAI reaches another milestone with new ChatGPT-4 Turbo

OpenAI has released the new GPT-4 Turbo model with the version number 2023-04-09. This is outstanding for two reasons.

• First, OpenAI has finally used sensible version numbers. It only took one year of progress.

• Second, this model beats all the other models on our LLM benchmarks. It takes the top place with a substantial gap to the second place.

This score jump comes from nearly perfect stores in CRM and Docs categories. Plus OpenAI has finally fixed the instruction following problem with few-shots, that was causing Integrate category to be so low.

GPT-4 Turbo 2023-04-09 is currently our default recommendation for the new LLM-driven projects that need the best performing LLM to get started.

Llama 3 70B and 8B

Meta has just released new models in its third generation. We have tested instruct versions of 70B and 8B for the usability in LLM-driven products.

Llama 3 70B had a bumpy start - the upload to HuggingFace had bugs with tokens in chat template processing. Once these were fixed, the model started working better, on the level of old generations of Anthropic Claude v2.

Note that we tested the b8 quantized model to properly fit 2xA100 80GB SMX cards. There is a possibility that f16 might give slightly better results.

Llama 3 8B Instruct performed much better on the benchmarks, pushing forward state of the art that is made available by Meta. This model has surprisingly good overall scores and a good “Reason” capability. There is a strong chance that product-oriented fine-tune of Llama3 8B Instruct would be able to push this model to the TOP-10.

Long-term trends

Now let's look at the bigger picture: Where is the industry heading with all this?

Further long-term LLM trends

• New functional capabilities of LLMs

  LLMs gain new functional capabilities that are not even covered in this benchmark: Function calls, multimodality, data grounding. The latest version of LLM Under the Hood expands on this theme.
   

• Experiments with new LLM architectures

  Companies also get bold and try to experiment with new LLM architectures outside the classical transformers' architecture. Mixture of Experts was popularised by Mistral, although many believe that GPT also uses it. Recurrent Neural Networks also experience a comeback as a way to solve context size limitations. For example: RWKV Language Model, Recurrent Gemma from Google Deep Mind (Griffin architecture).

Powerful models with low computing power

What is interesting about these models - they show decent capabilities, while requiring substantially less compute. For instance, we got a report of 0.4B version of RWKV running on a low-end Android phone with a tolerable speed (CPU-only inference).

Where are we heading with all this?

DEMOCRATIZATION OF AI

Expect the models to continue to get better, cheaper and more powerful. Sam Altman calls this the "democratization of AI". This applies to both cloud models and locally available models.

If you are in the process of building an LLM-driven system, expect that by the time the system is delivered, the underlying LLM will be much more powerful. In fact, you can take this into account and build a long-term strategy around it.

ADAPTABLE SYSTEMS

You can do that, for example, by designing LLM-driven systems to be transparent, auditable and capable of continuously adapting to the changing context.