Arcee AI unveils SuperNova: A customizable, instruction-adherent model for enterprises

Arcee AI launched SuperNova today, a 70 billion parameter language model designed for enterprise deployment, featuring advanced instruction-following capabilities and full customization options. The model aims to provide a powerful, ownable alternative to API-based services from OpenAI and Anthropic, addressing key concerns around data privacy, model stability and customization.

In an AI landscape dominated by cloud-based APIs, Arcee AI is taking a different approach with SuperNova. The large language model (LLM) can be deployed and customized within an enterprise’s own infrastructure. Released today, SuperNova is built on Meta’s Llama-3.1-70B-Instruct architecture and employs a novel post-training process that Arcee claims results in superior instruction adherence and adaptability to specific business needs.

Technical innovations

SuperNova’s development involved a multi-faceted approach to post-training, as explained by Lucas Atkins, lead engineer on the project:

“We trained three models at once. One was distilled from Llama 405B. Another was trained with a dataset we generated with our EvolKit repository. And the third was doing a pretty exhaustive DPO on top of the current Llama 3 instruct. At the end, we use a new kind of merging technique to combine all three, preserving the strengths of each one.”

This process, which Arcee considers proprietary, resulted in what they claim to be highly advanced instruction-following capabilities. The distillation from a 405B parameter model is particularly noteworthy, as it suggests that SuperNova may capture some of the capabilities of much larger models while remaining deployable on more modest hardware.

“As someone who tinkers with these models all day, both closed and open source, this one has been genuinely impressive to me,” Atkins added. “The big one here is instruction following, which was making it adhere very, very closely to the user or the organization’s needs.”

The use of EvolKit, Arcee’s synthetic data generation pipeline, is another key component of their approach. This tool, which will be open-sourced, allows for the creation of complex question-answer pairs that can be used to fine-tune models for specific tasks or domains. This could be particularly valuable for enterprises looking to adapt the model to their unique use cases.

Enterprise deployment and customization

SuperNova is designed to be deployed within an enterprise’s own cloud environment, starting with AWS Marketplace availability. Arcee is also working on making it available on Google and Azure marketplaces. Mark McQuade, co-founder of Arcee AI, highlighted the deployment process:

“The model gets deployed into your AWS VPC, but it also spins up a web server and a chat interface and a database to store your chat history. Everyone in your organization can interact with it.”

This deployment model addresses key enterprise concerns around data privacy and model stability. Unlike API-based services that can deprecate or change without notice, SuperNova provides businesses with full control over their AI assets. This is particularly relevant in light of recent events in the AI industry, as McQuade pointed out:

“OpenAI just deprecated 3.5… a lot of companies built up businesses around the API for 3.5. So that API changes, your app dies. In our world, nothing changes unless you change it, because it’s your model, your way to run it.”

The ability to deploy SuperNova within a company’s own Virtual Private Cloud (VPC) ensures that sensitive data never leaves the organization’s control. This can be important for companies in regulated industries or those dealing with confidential information.

Customization and continuous improvement

A key feature of SuperNova is its ability to be fine-tuned and retrained within the enterprise environment. Atkins explained the process and its benefits:

“Over time, we can retrain the model entirely within your own environment to better align with your preferences. As we save those chats, if you desire to have the model improve across the board for your unique preferences as a business, we have the ability to do that without ever having that data leave your system.”

This capability allows technical teams to adapt the model to specific domain knowledge or company-specific requirements over time. It’s a significant advantage over cloud-based API services, which typically don’t allow for this level of customization.

The continuous improvement aspect is particularly noteworthy. As the model interacts with users within an organization, it can learn from these interactions and improve its performance on company-specific tasks. This creates a virtuous cycle where the more the model is used, the more valuable it becomes to the organization.

Open source components

While the full 70B model isn’t open-source, Arcee is releasing several components for the developer community:

1. A free API for testing and evaluation: This allows developers to experiment with SuperNova without committing to a full deployment.
2. SuperNova-Lite: An 8B parameter open-source version of the model. This smaller model could be useful for developers working on resource-constrained environments or for those who want to understand the architecture before deploying the full model.
3. EvolKit: Their dataset generation pipeline for creating complex QA pairs. This tool could be valuable for organizations looking to create custom training data for their specific use cases.

By open-sourcing these components, Arcee is contributing to the broader AI community while also providing potential customers with tools to evaluate and customize their offering. Arcee SuperNova is also available in AWS Marketplace.

Performance claims and benchmarks

Arcee claims SuperNova performs well in various areas, with a particular strength in mathematical reasoning. “This one is pretty outstanding on math benchmarks,” Atkins noted. However, the company is encouraging third-party evaluations to verify their claims.

“We’re going to have an API available for people to hit. And if there are third-parties that want to run credible benchmarking to evaluate it themselves, we can make arrangements to provide them with access to the weights. We want to have full transparency with this model” Atkins said.

This openness to third-party evaluation is commendable, as it allows for independent verification of Arcee’s claims. It will be particularly interesting to see how SuperNova performs on standard benchmarks compared to models from OpenAI, Anthropic and other leading AI companies.

The emphasis on mathematical reasoning is noteworthy, as this has been a challenging area for many language models. If SuperNova indeed excels in this domain, it could be particularly valuable for industries such as finance, engineering and scientific research.

Implications for Enterprise AI strategy

The release of SuperNova comes at a time when many enterprises are reevaluating their AI strategies. While cloud-based API services have dominated the landscape, there’s growing interest in deployable, customizable models that offer more control and flexibility.

SuperNova’s approach addresses several key concerns:

1. Data Privacy: By deploying within a company’s own infrastructure, SuperNova ensures that sensitive data never leaves the organization’s control.
2. Model Stability: Unlike API services that can change or deprecate without notice, SuperNova provides a stable base that only changes when the organization chooses to update it.
3. Customization: The ability to fine-tune and retrain the model on company-specific data allows for deep customization that isn’t possible with most API services.
4. Cost Control: While initial deployment may require significant resources, the long-term cost of running SuperNova could be lower than paying for API calls at scale.
5. Competitive Advantage: A customized, continuously improving AI model could provide significant competitive advantages in industries where AI-driven insights are critical.

The AI sovereignty dilemma

As enterprises navigate the rapidly evolving AI landscape, SuperNova’s release reveals a growing tension in the industry: the trade-off between the convenience and power of cloud-based AI services and the control and customization offered by deployable models. This dichotomy presents what we might call the “AI Sovereignty Dilemma.”

On one side, cloud-based API services like GPT-4 and Claude offer state-of-the-art performance and constant updates, but at the cost of data privacy concerns and limited customization. On the other, models like SuperNova promise full control and customization but require significant in-house expertise to deploy and maintain.

Arcee’s approach with SuperNova attempts to bridge this gap, offering a model that can be deployed on-premise while still providing capabilities that aim to rival leading cloud-based services. This hybrid approach could be particularly appealing to industries with strict regulatory requirements or those dealing with highly sensitive data.

However, the success of this model will depend on several factors:

1. Performance Parity: Can models like SuperNova truly match the capabilities of constantly updated cloud models?
2. Ease of Deployment: Will enterprises find the deployment and maintenance process manageable?
3. Customization Benefits: Will the ability to fine-tune the model on proprietary data provide a significant competitive advantage?
4. Cost-Effectiveness: Over time, will the total cost of ownership for models like SuperNova be lower than using cloud-based APIs at scale?

The release of SuperNova signals a potential shift in the enterprise AI landscape. It challenges the notion that state-of-the-art AI capabilities are only accessible through cloud APIs and pushes back against the centralization of AI power in the hands of a few tech giants.

SuperNova and similar models represent a new chapter in the enterprise AI story. They offer a vision of AI that is more controllable, customizable and aligned with specific business needs. Whether this vision will supplant or complement the current cloud-dominated paradigm remains to be seen, but one thing is clear: the battle for the future of enterprise AI is intensifying, and models like SuperNova are at the forefront of this revolution.