OpenTSLM: Language Models That Understand Time-Series (Stanford, ETH, Google) (www.opentslm.com)

OpenTSLM introduces Time-Series Language Models (TSLMs), a new class of multimodal foundation models that treat time series (heartbeat, price ticks, sensor pulses, logs, clicks) as a native modality alongside text. The project, led by researchers from Stanford, ETH and collaborators from industry, has released lightweight open base models trained on public data and outlines a “frontier” tier of proprietary models trained on specialized datasets for enterprise APIs and fine-tuning. Their results claim order-of-magnitude gains in temporal reasoning while running on smaller, faster backbones, enabling direct natural-language reasoning, explanation, and forecasting over temporal data rather than treating time series as an external add-on. This is significant because most current LLMs don’t natively model continuous temporal structure; TSLMs aim to be the temporal interface that connects real‑world signals to intelligent decisions and autonomous agents. Practically, that could accelerate use cases in proactive healthcare (continuous monitoring and alerting), adaptive robotics, resilient infrastructure management, and time-aware analytics. The OpenTSLM open-core + frontier-edge strategy also signals an ecosystem play: openly available base models to set standards and foster research, with proprietary variants offering production-grade performance and domain specialization.