Maestro works through Cadence's AI Super Agents, which interpret design intent, break tasks down and orchestrate the Maestro models alongside other frontier and open-weight models, including Nvidia's Nemotron. Cadence also uses reinforcement learning, with its deterministic solvers judging outputs against engineering objectives and feeding the results back into the models.

Flexibility is a selling point: customers can run the models on Cadence's cloud, public clouds or their own on-premises environments, and post-train them on their own proprietary data and workflows using Cadence's fine-tuning tools, potentially on their own GPUs.

Maestro lands less than a month after rival Synopsys announced its own specialised model, GPT-Synopsys, and the two EDA giants split on where customer design data goes: Cadence lets customers post-train its models on their own data and infrastructure, while GPT-Synopsys runs on OpenAI-hosted infrastructure without using customer data for training.

The comparison with general-purpose models matters for cost as well as quality. Cadence executive Paul Cunningham said agent costs have "really started to explode", and Maestro use counts toward Cadence's consumption-based agent subscription, though he did not say whether the token savings would cut customers' bills. One caveat: Cadence did not specify what its 30% quality figure measures. In chip design, quality of results usually refers to power, performance and area, and both claimed figures are benchmarks against the unnamed open-source foundation model Maestro started from.