SemiKong, built by Aitomatic and other collaborators under the Foundation Models workgroup of the AI Alliance, serves as a shared resource reducing development costs for companies across the sector for applications spanning chip design, manufacturing, and testing.

“The semiconductor industry lacked a specialized, open source language model tailored to its unique terminology, processes, and knowledge base," says Christopher Nguyen, CEO of Aitomatic. “SemiKong closes the gap, promoting innovation and collaboration across the industry to adopt AI in accelerating many mission-critical manufacturing and operational procedures. We believe that the future of the semiconductor industry lies in the convergence of deep domain expertise and advanced AI. Projects like SemiKong are just the beginning.”

The team used Llama 3.1 70B as the base model and extensively fine-tuned it with semiconductor industry documents, research papers, and anonymized design and manufacturing data.

“The joint effort employing an open source approach allowed the team to pool resources and expertise, accelerating development,” Nguyen says, noting that SemiKong outperforms several state-of-the-art generic closed-source LLMs in understanding and generating semiconductor-related content.

Expected benefits include a 20-30% reduction in time-to-market for new chip designs and a 15-25% improvement in first-time-right rates in chip manufacturing. SemiKong is also expected to speed up the onboarding of new semiconductor professionals by 40-50%.

In the semiconductor industry, human expertise is central. Many critical problems require deep experience built over decades and thousands of decisions. But as veteran experts are retiring faster than their knowledge is replaced, there is an unprecedented challenge—and opportunity—for AI to preserve and scale their expertise.

Aitomatic’s Domain-Expert Agents (DXAs), based on a neurosymbolic agentic-AI architecture named DANA, address the expertise bottleneck by a systematic three-phase Capture-Train-Apply agent lifecycle. First, expert knowledge and programs are organized and structured. Next, that human expertise is augmented with synthetic knowledge to train a DXA to handle a wide variety of scenarios. The trained DXA is then connected to the semiconductor companies’ manufacturing execution systems and IoT platforms to automate key technical analyses and decisions.

Semiconductor DXAs’ inner workings are powered by Llama: SemiKong, based on Llama 3.1 70B, typically plays the role of a central neurosymbolic planning-and-reasoning brain, while lightweight Llama 3.2 1B and 3B models facilitate fast access to diverse informational resources such as IoT data streams and technical documentation.

DXAs have proven highly effective in real-world equipment manufacturing and troubleshooting at leading semiconductor equipment makers and integrated-circuit (IC) design companies. Applications range from process optimization to manufacturing yield monitoring and predictive maintenance. For instance, an intricate and hours-long analytical workflow like etching recipe formulation, which involves dozens of machine parameters that affect key physical operations and chemical reactions, can now be performed in minutes when process engineers are assisted by DXAs. In another high-impact use case, veteran field engineers’ design and product-operation knowledge is captured into DXAs that are deployed to advise less experienced field engineers globally, thereby improving end-customer support accuracy and speed while significantly reducing the organization’s dependency on a dozen senior experts.

“Lightweight models available from the 3.2 release are a strong differentiating factor setting Llama apart from other open source LLM ecosystems,” says Vinh Luong, Head of Open Source at Aitomatic and lead engineer of the company’s OpenSSA framework for small specialist agents. “They allow us to construct and deploy efficient DXAs in resource-constrained manufacturing environments to perform fast sensor data analysis, parameter calibration and production issue detection.”

Nguyen notes that the open source approach enables companies like Aitomatic to deliver cutting-edge AI solutions more rapidly and cost-effectively.

“Llama’s permissive license supports both industry-wide collaboration and individual company customization,” Nguyen says, enabling companies to develop specialized models and agents more flexibly and cost-effectively than starting from scratch.

For smaller companies and institutions, it democratizes access to advanced AI technologies, allowing them to compete and innovate in specialized fields like semiconductor design without massive R&D budgets. Semiconductor companies are able to maintain the technical independence needed to fully own and control key manufacturing and operational assets.

“We are thrilled to see early transformative outcomes from SemiKong and DXAs in the semiconductor field, and optimistic about AI driving continued improvements in equipment uptime, production yield and engineers’ productivity throughout this critical industry,” says Luong.