Ai21 Together Ai Partnership

AI21 × Together AI: Open-Source AI for the Enterprise

Skip to Main Menu

Skip to Main Content

Skip to Footer

Back to Blog

-->

Back to Blog

Open-source development has reshaped AI research, but enterprise adoption still lags behind. More than 75% of organizations say they plan to expand their use of open-source AI, with many expecting those choices to reduce implementation costs. Yet, only 13% of AI workloads today actually run on open-source models. The gap is real, and it’s one enterprises feel every time they try to operationalize open-source systems at scale.

This partnership exists to change that.

Today, AI21 Labs and Together AI are bringing their strengths together to make open-source AI practical for enterprise use. Our integration combines AI21 Maestro, our orchestration system for building reliable, multi-step AI agents, with Together AI’s high-performance open-source model platform — including native access to the full Jamba open-source family. This gives enterprises a unified, efficient way to evaluate, route, and operationalize diverse models without restructuring their workflows.

Why this matters

As generative AI shifts from experimentation to core business operations, model choice becomes foundational. Organizations need visibility into how models behave, the ability to compare options side-by-side, and the freedom to optimize for accuracy, cost, or performance depending on the use case.

Together AI provides a powerful ecosystem of optimized open-source models. AI21 Maestro provides the orchestration layer that makes those models usable in mission-critical environments.

In the words of Vipul Ved Prakash, CEO of Together AI:

“Open-source is the way forward for AI as a platform. We believe open models democratize frontier capabilities, making them accessible to any organization and offering more control and cost efficiencies. Our collaboration with AI21 brings those models natively into enterprise environments in a seamless, actionable way.”

How model choice works inside AI21 Maestro

To show how this works in practice, here’s a quick walkthrough of the model-selection experience inside AI21 Maestro:

In this demo, teams browse available models, see capability summaries, and assign them to steps within agentic workflows. They can use a Jamba model, switch to an open-source LLM from Together AI, or combine multiple models depending on the task. All of this happens without rewriting pipelines or re-architecting infrastructure.

Our CEO, Ori Goshen, put it like this:

“Maestro was built to orchestrate models intelligently, regardless of origin. By integrating Together AI’s open-source ecosystem, we’re enabling enterprises to aggregate the intelligence of a diverse portfolio of open-source models – providing true model choice, full transparency, and the building blocks to operationalize agent systems reliably.”

A new open architecture for the enterprise

Over the past year, enterprises have made it clear: they want AI systems they can inspect, govern, and adapt with confidence. This integration delivers exactly that.

Together AI provides a flexible, high-performance open-source model ecosystem.

AI21 Maestro operationalizes those models with orchestration, governance, and reliable execution across complex workflows.

The Jamba open-source family adds efficient long-context reasoning built for enterprise-scale tasks.

With this partnership, enterprises no longer face a tradeoff between openness and operational rigor. They get the transparency and control of open-source models alongside the reliability and governance required for production.

This is what real model choice looks like. And it’s now available natively through AI21 Maestro and Together AI.

Discover more

Jun 25, 2026

Token spend isn’t going down. You need more than naive routing to manage it

Labs in Front

-->

Jun 24, 2026

Tipping the scales: Merging weak agents into a state-of-the-art deep researcher

Labs in Front

-->

Jun 4, 2026

First scale, then enrich: How the right execution strategy helped us reach state-of-the-art on SWE-rebench