Goodfire raises $150M Series B at $1.25B for model interpretability

B Capital led the round. Existing investors Juniper Ventures, Menlo Ventures, Lightspeed, South Park Commons and Wing Venture Capital took part. New backers include DFJ Growth, Salesforce Ventures and Eric Schmidt. The money is earmarked for frontier research, the next version of the core product, and partnerships in AI agents and life sciences. The planned platform would let users find the parts of a model responsible for a behavior, then train or intervene on those parts directly. The company has early proof points outside language models. Applying its methods to an epigenetic model from Prima Mente, it identified a new class of Alzheimer's biomarkers. It also says one real-world project cut hallucinations by about half. Staff include Tom McGrath, who founded the interpretability team at Google DeepMind, and Nick Cammarata, a core contributor to OpenAI's interpretability team. For founders, the open question is demand. One analysis notes it is unclear whether buyers will pay for internal model audits before regulation or failures force the issue. Teams shipping models into regulated or high-stakes settings should expect inspection of model internals to become a vendor category. Teams that fine-tune models should watch whether targeted interventions beat retraining on cost and side effects.