Zeta Global (NYSE: ZETA) yesterday set out plans for the Athena Inference Model, a collection of specialized open-weight models that the company built with Fireworks on top of NVIDIA's Nemotron open models, according to Zeta. The models are still in development, and the company said they will join the AthenaOS beta later this year.

In Short

Zeta Global, a marketing and data software company, said yesterday that it is building its own set of smaller AI models, tuned for its AthenaOS platform, on top of openly available models from NVIDIA. This matters to businesses that feed customer data into AI tools, because Zeta says the models are meant to let companies keep control of that data while still getting recommendations. The models are not finished yet, they are due in a test version of AthenaOS later this year, and the accuracy claim comes from Zeta alone, with no benchmark published.

What Zeta described

Zeta's account of the Athena Inference Model, shortened to AIM, fits in a handful of paragraphs and a single status line. The company describes "a collection of specialized open-weight models powering specific AthenaOS use cases", developed with Fireworks and powered by NVIDIA Nemotron open models. AthenaOS, in Zeta's words, is "the enterprise intelligence operating system that dynamically builds around a user's objective and brings intelligence into the tools and applications where work already happens."

According to Zeta, AIM is designed for tasks that include "interpreting domain language and following brand voice, formatting and output requirements". It will work alongside the company's existing predictive models, with the stated aim of helping AthenaOS produce "more relevant intelligence and better recommendations". Each enterprise's objectives are meant to guide the resulting action, though Zeta gave no mechanism for how those objectives reach the models.

Status is the plainest part. AIM is "currently in development and will be part of the AthenaOS beta later this year", Zeta said. No beta start date appeared in the company's material, nor did the number of models in the collection, parameter counts, pricing or customer names. Which Nemotron model sits underneath was also left out.

Performance claims are thin. "In early testing, AIM achieved accuracy on par with state-of-the-art industry leading models," Zeta said. No benchmark, task set or comparison model is named, so the figure is vendor-supplied and cannot be checked from the material. Benny Chen, Co-Founder at Fireworks, went further, saying the combination of Zeta's proprietary intelligence and Fireworks' infrastructure makes the models more accurate and more efficient and tailored to Zeta's business needs - again without a stated baseline. Early results, in other words, rest on Zeta's word alone!

The executive framing is consistent across the three statements. "Models are increasingly available, but context is owned," said David A. Steinberg, Zeta's Co-Founder, Chairman and CEO, who added that AIM helps organizations "harness their most strategic assets securely and effectively". Nate Yohannes, President of Data & AI Lab and Global Head of Research and Development, took a similar line: "General-purpose models provide the foundation, but open models unlock the proprietary data and specialization that drive the next leap in performance." Chen's contribution was that "data sensitivity cannot be an afterthought" for enterprises. The common thread treats general-purpose models as a base layer and places differentiation in proprietary data.

Open weights, Nemotron and Fireworks

What the Nemotron family supplies

The term open-weight describes a model whose trained parameters are published for others to download, run and adapt. The term is narrower than open source, because training data and code can stay private. NVIDIA's Nemotron 3 release goes beyond weights: according to NVIDIA, the family released on December 15, 2025 ships with about 3 trillion tokens of pretraining, post-training and reinforcement-learning datasets, plus the NeMo Gym and NeMo RL libraries and the NeMo Evaluator tool.

The lineup comes in three sizes, per NVIDIA. Nano has roughly 30 billion total parameters with up to 3 billion active per token and a 1-million-token context window. Super has roughly 100 billion with up to 10 billion active, and Ultra roughly 500 billion with up to 50 billion active. The family relies on a mixture-of-experts design, in which only part of the network runs for each token. Nano was available at launch; Super and Ultra were expected in the first half of 2026.

Which of the three underlies AIM? Zeta does not say. The question carries weight because serving cost and latency track the number of active parameters more closely than the total, and a collection of task-specific models could in principle draw on more than one size.

What Fireworks supplies

Fireworks runs an inference and training platform for open-weight and custom models, according to TCV's profile of the company, and lets enterprises fine-tune those models on their own data and serve them in production. The company was founded in 2022 and is based in San Mateo, California. Bloomberg reported on May 27, 2026 that Fireworks was in talks about funding at a valuation of $15 billion; the report described talks rather than a completed round.

Zeta said AIM is "fine-tuned using Fireworks infrastructure". Chen described Fireworks' role as "infrastructure for training, tuning, and serving open models". Tuning, then, is placed with the third-party platform. Whether the production serving of AIM for Zeta customers also runs there is a point the material leaves open, as is the tuning method and what data was used. The release refers to "Zeta's proprietary intelligence" without defining it.

Data control: the claim and what is missing

The data-control language is the part most relevant to marketers, whose customer files are the raw material for any such system. According to Zeta, AIM supports "enterprise data-control and sovereignty requirements", letting customers apply their first-party data "in a governed environment while retaining control over how it is used".

Left undefined: what makes an environment governed, whether the models run inside a customer's cloud account, on Zeta's infrastructure or at Fireworks, and whether any customer data enters training. No certifications, audit rights or deployment options are listed. Steinberg tied AIM to the effort to "advance data and AI sovereignty", yet sovereignty in the regulatory sense usually turns on location and jurisdiction, and neither is addressed.

Other Zeta statements give partial context. On the August 4, 2026 earnings call, according to Nasdaq's summary, Steinberg said no large language models have access to data inside the Zeta Data Cloud, that OpenAI powers Athena's voice functionality, and that Zeta's own inference models make the decisions using the Data Cloud. The Palantir arrangement is a separate matter: PPC Land reported that Foundry components can be deployed inside a customer's existing infrastructure, so data need not move to a third-party cloud. Nothing in yesterday's material connects AIM to either arrangement.

Where AIM sits in the Athena build-out

Athena's model history is short and already layered. Zeta showed Athena at Zeta Live on October 9, 2025, set out a collaboration with OpenAI on January 5, 2026, and made Athena generally available on March 24, 2026, when it was built on OpenAI models and Zeta still called itself an AI Marketing Cloud company. On June 23, 2026, Zeta and Palantir disclosed a partnership under which the Data Cloud is to be rebuilt on Palantir's Foundry platform, with Athena as the intelligence layer on top.

The wording has moved too. Yesterday's company description calls Zeta "the intelligent AI infrastructure company", which differs from the March self-description. Zeta's proprietary Data Cloud and identity graph remain the stated foundation.

Usage figures give a sense of the volume such a stack has to carry. Per Nasdaq's summary of the August 4 call, about 130 days after availability more than 40% of super-scaled customers, defined there as those with at least $1 million in annual revenue, had become monthly active users of Athena, and 83% of customer interactions happened by voice. On the same date, according to Zeta, second-quarter revenue reached $443 million, up 44% (28% excluding the effect of acquisitions), with GAAP net income of $8 million. Acquisitions have contributed to that growth, among them Marigold's enterprise business, agreed on September 30, 2025. Zeta's list of recent news also carries an agreement to acquire Digital Audience, an AthenaOS item tied to Zeta Live 2026, and a planned discussion between Palantir's Alex Karp and Steinberg on data sovereignty and enterprise AI.

AIM would add an open-weight tier to a stack that, by Zeta's own account on the call, already used Zeta-built inference models for decisions and OpenAI for voice. Yesterday's material does not say whether AIM displaces any component built on OpenAI models or sits only beside Zeta's predictive models, as stated.

Context for the marketing industry

No product exists to test yet, so the significance lies in positioning. Zeta's thesis is that model access is spreading while context stays proprietary. The sector has been moving in a similar direction: Publicis's $2.5 billion purchase of LiveRamp, covered by PPC Land on May 17, 2026, also turned on data infrastructure for agentic AI. Whether specialized open-weight models tuned by a platform vendor outperform general-purpose models on marketing tasks is the empirical question, and Zeta has not yet supplied the means to answer it.

Items the material leaves open:

  • the Nemotron variant and the number of models in AIM
  • the benchmark, task set and comparison models behind the "on par" accuracy claim
  • the hosting location and the meaning of "governed environment"
  • whether customer data enters training or tuning
  • the beta start date, access terms and pricing

Timeline

Summary

Who: Zeta Global (NYSE: ZETA), the marketing and data software company led by CEO David A. Steinberg, with Fireworks, an inference and training platform, and NVIDIA's Nemotron open models as the base.

What: The Athena Inference Model (AIM), a collection of specialized open-weight models for AthenaOS tasks such as interpreting domain language and following brand voice, formatting and output requirements, fine-tuned with Fireworks infrastructure and set to work alongside Zeta's existing predictive models.

When: Yesterday. The models are in development and are due to join the AthenaOS beta later this year.

Where: Inside AthenaOS, Zeta's enterprise platform; Zeta is headquartered in New York City. Where the models are hosted and where customer data is processed is not stated.

Why: According to Zeta, to help customers generate insights faster, improve efficiency, scale specialized AI across the enterprise and meet data-control and sovereignty requirements, with accuracy it describes as on par with leading models in early testing.