Expanding Horizons: Embracing the Multi-Model Future with Martian

May 2, 2024

•

5 min read

Etan Ginsberg

Summary · TLDR

The AI industry is racing towards a multi-model future.
Technological advancements, demand for specialization, and unique data sources drive this shift.
Last month, over 20 new language models were released, marking a record high; and HuggingFace surpassed hosting 1 million ML models.
Martian's platform has added a variety of these new models such as Llama-3, DBRX, and Claude-3, enhancing capabilities across different applications.
A multi-model world gives Martian more opportunities to achieve better performance customization and cost efficiency.

‍

The world AI is evolving at an unprecedented pace. Last month alone saw the release of more than 20 frontier language models (LLMs) – the highest number ever in a single month. Meanwhile, HuggingFace, a leading platform for machine learning models, has crossed the incredible milestone of hosting over 1 million general ML models. This explosion in the number and variety of AI models is a clear indication that we are heading towards a multi-model future.

The Rise of the Multi-Model Paradigm

The AI industry is rapidly shifting from a single-model paradigm to a multi-model one. This shift is driven by several factors:

Technological Advancements: The cost of training AI models is decreasing fast due to advancements in hardware, algorithms, and infrastructure. This makes it easier and more affordable to create a wide variety of specialized models.
Demand for Specialization: There is a growing demand for models that are fine-tuned for specific tasks and industries, as AI is applied to more domains. A one-size-fits-all approach is no longer sufficient.
Unique Data Sources: The performance of AI models is heavily dependent on the quality and relevance of the data they are trained on. Organizations with unique data sources can create models with unique competencies that outperform generic models.

The biggest players in the AI industry, including Google, Microsoft, Amazon, and OpenAI, are all betting on a multi-model future. Microsoft Azure now supports diverse open-source models like Llama and Mistral, while Amazon has launched Amazon Bedrock for streamlined model management. OpenAI has introduced GPT-Store, expanding access to various GPT models. These developments highlight their commitment to providing tools and platforms for creating, hosting, and deploying a wide variety of models.

Martian: Your Guide in the Multi-Model World

At Martian, we are committed to helping our users navigate and thrive in this multi-model world. Our platform continuously integrates the latest and greatest models, ensuring that our users always have access to the best tools for their needs.

We have added a wide range of new models to our platform in the last month, including:

Llama-3 by Meta: Meta's biggest release yet, open-sourcing an 8B parameter model outperforming GPT-3.5 on many benchmarks and a 70B parameter model approaching GPT-4 level performance on some tasks.
DBRX by Databricks: Leveraging a fine-grained mixture-of-experts (MoE) architecture, DBRX achieves 2x faster inference than LLaMA2-70B, 40% smaller size than Grok-1, and up to 150 tok/s/user generation speed when hosted on Mosaic AI Model Serving. This robust model is optimized for data extraction and analysis.
Claude-3 models by Anthropic: The Claude 3 family includes Opus, leading in general intelligence at the level of GPT-4; Sonnet, offering 2x speed over Claude 2 for most workloads; and Haiku, providing unmatched speed. All models can handle 200K+ token context windows, with Opus doubling accuracy on complex questions while reducing hallucinations.
Mistral Large by Mistral AI: This powerful multilingual model achieves top-tier reasoning and strong performance on benchmarks. Natively fluent in English, French, Spanish, German, and Italian, Mistral Large has a 32K token context window and excels at coding and math tasks.
Online models by Perplexity AI and You.com: These models focus on obtaining factual, up-to-date responses by leveraging web data. Models like pplx-7b and pplx-70b surpass offline models in providing current, real-world information.

Unleashing the Power of Multiple Models

The addition of these models to our platform significantly expands the capabilities and use cases that our users can address. From basic automation tasks to complex problem-solving, our model lineup offers a tool for every job.

What sets Martian apart is not just the diversity of models but the simplicity of integration. Our intelligent routing system seamlessly selects the optimal model for each request, ensuring peak performance and cost-effectiveness with only a minimal two-line code change to integrate —matching the OpenAI format. This ease of integration is crucial because traditionally, each API provider uses a unique syntax, making the use of different models a cumbersome and error-prone process.

Getting ready for the Multi-model world

The shift towards a multi-model paradigm represents a significant opportunity for businesses and developers alike. By embracing this shift, you can unlock new capabilities, improve your operations, and drive innovation in your field.

We invite you to explore the new models available on our platform, experiment with different approaches, and share your feedback with us. For more information and technical support, please visit our documentation site or reach out to our support team. For more detailed information on each model and for technical support, please visit our resources page or contact us directly at contact@withmartian.com.

The future of AI is multi-model, and with Martian, you're ready for it.

‍

Thank you! Your submission has been received!

Oops! Something went wrong while submitting the form.

Etan Ginsberg