Cadence Design Systems on October 26, 2026 announced Maestro, a portfolio of domain-specialized AI models meant to power the company's AI Super Agents for semiconductor and system design. Cadence says the first Maestro models will reach select early-access customers in the fourth quarter of 2026; it has not given a general-availability date or published pricing for the models on their own.
According to Cadence's announcement, Maestro is not a single model but a mix of post-trained large language models in several sizes and architectures, trained decision models, and what the company calls self-improving agent skills. The models sit underneath Cadence's four AI Super Agents, AuraStack, ChipStack, InnoStack and ViraStack, which Cadence says interpret a designer's high-level objective, break it into tasks, and decide when to call Maestro, outside frontier or open-weight models, or Cadence's own electronic design automation (EDA) and system design automation (SDA) tools.
Nvidia is the main technology partner. Cadence and Nvidia say the portfolio incorporates Nvidia's open-weight Nemotron models, and the two companies are working on agent security through Nvidia's OpenShell reference architecture. Tim Costa, who runs Nvidia's industrial and computational engineering business, said in the release that grounding reinforcement learning in feedback from EDA tools and deterministic solvers can teach agents to make better design decisions.
That feedback loop is the core of Cadence's pitch. The company says its solvers check model outputs against engineering objectives, constraints and ground truth, and that those verified results feed back into reinforcement-learning training. Cadence also says balancing conventional tool calls against language-model calls cuts token use, a point senior vice president Paul Cunningham underscored to Tom's Hardware when he said agent costs have "really started to explode."
Cadence's headline numbers are its own and come with caveats. The company claims initial design-workflow results of up to 30% higher quality of results and up to five times better token efficiency than a general-purpose foundation model. Tom's Hardware reports that Cadence did not specify what the 30% figure measures, and that the comparison baseline is the unnamed open-source model Maestro was trained from rather than a leading commercial system. Cadence has not released test data supporting either figure, and DigiEditorial has not independently verified them.
Deployment is where Cadence is drawing the sharpest contrast with its main rival. Cadence says customers can run the full agent stack on its own cloud, on public clouds, or entirely on premises, and can post-train Maestro models on their own proprietary designs and workflows. Reuters notes that this differs from Synopsys, which said on September 30 that it would build a model called GPT-Synopsys with OpenAI that runs on OpenAI-hosted infrastructure. Cunningham told Reuters that Cadence is pursuing open- and closed-weight approaches "equally" and that working with commercial labs is part of its plans, without giving details.
Two customers, Socionext and Silicon Labs, are quoted in the announcement, though both describe results from Cadence's existing agents rather than from Maestro itself. Socionext said it is deploying the InnoStack agent in production physical design and expects up to fivefold productivity gains, and Silicon Labs said the ChipStack agent sped up debugging in pre-silicon low-power verification. Tom's Hardware reports that Maestro usage will count toward Cadence's agent subscription, which includes a consumption allowance with charges beyond it.
DigiEditorial treats Cadence's October 26 press release as the primary source and corroborated the announcement date, Nvidia's role, deployment options and early-access timing against Reuters' report and Tom's Hardware's interviews with Cadence executives. Performance figures attributed to Cadence are company claims from initial results, and the models are not yet broadly available.