On the (re)-prioritization of open-source AI

Pinterest shares its strategic shift toward fine-tuned open-source AI models, achieving comparable performance at less than 10% the cost of proprietary models.

• •Pinterest categorizes its AI investments into three modalities: user recommendation systems (built in-house), visual encoders/diffusion models (trained from scratch), and LLMs/VLMs (increasingly using open-source).
• •Open-source multimodal LLM architectures have leveled the playing field, shifting competitive differentiation to domain-specific data, fine-tuning, and product integration.
• •Pinterest Assistant uses a two-layer architecture: Pinterest-native multimodal retrieval/recommendation tools, plus a core multimodal LLM acting as an intelligent agentic router.
• •Key advantages of open-source adoption include order-of-magnitude inference cost reduction for image-heavy workloads, better personalization via native embedding integration, and more efficient long visual context processing.
• •The strategy mirrors the pre-LLM Alex