Key Takeaways: Local inference and model routing can help manage cost and protect sensitive data, but chip architects must account for memory capacity, sustained performance, changing models, and deployment security. For EDA, model quality depends on clean, trusted design data and integration into workflows, not just on choosing a larger model or fine-tuning its weights. Open weights can reduce dependence on proprietary AI services, but they do not automatically provide open training data, source code, or unrestricted usage rights. Open-weight AI models are creating new challenges for chip architects as rapidly changing workloads strain memory, interconnects, and compute flexibility. The focus is no longer just peak TOPS. The checklist has grown to include the best way to feed memory-hungry models, sustain performance across long-context and multi-agent inference, support rapidly changing numeric formats and sparse computation, and isolate models, data, and tools as part of the security boundary. Model architectures can change in weeks while silicon takes years, so the winners will build adaptable hardware-software platforms, rather than accelerators optimized for a benchmark that may be obsolete before tape-out. Massive data centers are not a prerequisite for AI-driven scientific advances. Science has different infrastructure needs from consumer AI platforms serving millions of users. Open-weight AI models — whose trained parameters are publicly available — can be deployed locally, potentially widening access, improving compute efficiency, and supporting independent research. Open weights also raise questions about control, cost, and who can build on AI technology. Sharad Chole, chief scientist and co-founder at Expedera , said publishing weights can reduce vendor lock-in as it allows users to host a model themselves or choose a service provider. “If you want to use a service, or host it yourself, there is no vendor lock-in in terms of who is providing the next token for you or who is servicing your next request. But there are geopolitical and investment sides to it. There is also a push for open hardware, like what Meta and others do to enable different competing solution providers. As of September 2026, the larger open-weight models come primarily from outside the U.S., either Europe or Asia, and we are moving toward trillion-parameter models.” Open-weight models make trained parameters downloadable, letting users run and adapt them on their own hardware rather than relying on a provider’s API, according to Guillaume Boillet, vice president of strategic marketing at Arteris . “They are not necessarily open source: training data, code and usage rights may remain restricted.” While open-weight AI model releases publish trained parameters, open-source releases aim to publish the full recipe, and closed models publish neither, said Ashish Darbari, CEO of Axiomise . Representative open-weight releases include Meta Llama, Alibaba Qwen, DeepSeek, Mistral, Google Gemma, Microsoft Phi, and OpenAI’s gpt-oss family. OpenAI reports that gpt-oss-120b runs on a single 80 GB GPU and gpt-oss-20b can run on devices with 16 GB of memory; both use MXFP4 quantization, Darbari noted. “These specifications illustrate the widening range of deployment options, but memory fit alone does not establish useful performance.” Some U.S.-based AI companies use knowledge distillation to develop smaller models from larger teacher models, but approaches vary. Smaller models may also be pretrained independently or developed through a mix of techniques. Large-scale pretraining remains computationally intensive, while distillation and post-training are important parts of the model-development process. Chole said the largest models continue to push toward trillion-parameter scales, while models intended for devices or local AI are often distilled into smaller versions. “It’s not like suddenly we’re going to have 100-billion-parameter models breaking all the benc