Not Just LLMs: Google’s New Specialized Models for Raw Sensor Data and Tabular Tables

Large language models are incredible at text, reasoning, and general knowledge. But when it comes to raw numerical data — especially high-volume sensor streams or large structured tables — they are often the wrong tool for the job. Dumping a year of heart-rate readings or a 100,000-row spreadsheet into ChatGPT turns precise numbers and relationships into a long “story.” The model loses structural information, struggles to truly “feel” numerical patterns, and quickly runs into context limits.

Google Research has just released two specialized foundation models that directly address these weaknesses: SensorFM for wearable sensor data and TabFM for tabular data. Both were introduced about two weeks ago and already look extremely promising.

---

SensorFM: Learning Directly from Raw Wearable Signals

SensorFM is trained directly on raw multimodal sensor data from wearables — no text conversion required. It was pre-trained on over one trillion minutes (roughly two billion hours) of data collected from five million consented participants across more than 100 countries and many Fitbit and Pixel Watch models.

The model ingests signals such as:

• Photoplethysmography (PPG) → heart rate and heart-rate variability;
• Accelerometry → movement and steps;
• Electrodermal activity (EDA);
• Skin temperature;
• Altimetry.

Crucially, it handles missing data natively. If someone takes off their watch, the model doesn’t break — real-world gaps are treated as a normal part of the signal during training.

Performance highlights:

• On 35 discriminative health tasks (cardiovascular, metabolic, sleep, demographics, lifestyle, and mental health), the frozen encoder + simple linear probe outperformed hand-engineered expert baselines on 34 out of 35 tasks.
• It particularly shines at detecting subtle signals that are hard to spot from pulse data alone — including depression and anxiety.
• Scaling experiments show consistent gains with more data and larger models, with no signs of saturation yet.

Instead of forcing an LLM to interpret noisy time-series data as text, SensorFM learns a general-purpose representation of human physiology. It can then be adapted efficiently (even with small amounts of labeled data) or used to ground a Personal Health Agent that generates clinically useful insights.

---

TabFM: Zero-Shot Foundation Model for Tabular Data

Tabular data is the lifeblood of business analytics, yet it has always been awkward for LLMs.

Tables have complex row/column relationships, mixed data types, and missing values. Traditional pipelines require data scientists to spend days cleaning data, engineering features, and tuning models like XGBoost.

TabFM flips this script. You can feed it a table as-is and get high-quality predictions in a single forward pass — zero-shot, with no training or hyperparameter tuning required.

Key details:

• Trained entirely on hundreds of millions of synthetic tables generated via structural causal models. This gave the model exposure to diverse distributions and complex feature relationships without needing massive real-world datasets.
• Uses a hybrid architecture with alternating row and column attention, plus row compression, specifically designed for two-dimensional tabular structure.
• On a broad benchmark (TabArena) covering dozens of classification and regression datasets, TabFM matches or beats heavily tuned classical models.
• Coming soon to Google BigQuery — you’ll be able to run advanced predictions with a simple `AI.PREDICT` SQL command.

This is a game-changer for analysts and data teams who don’t want (or need) to become machine learning experts just to get good predictions from their tables.

---

Why Specialized Models Beat “Just Throw It at an LLM”

LLMs process everything as text.

When you convert sensor time-series or large numerical tables into tokens:

• Fine-grained numerical relationships and ordering get diluted.
• The model treats data more like narrative than precise measurements.
• Context windows fill up extremely fast with high-volume data.

Specialized models like SensorFM and TabFM are trained on the actual modality — raw signals or native tabular structure.

They preserve what matters and scale efficiently.

---

The Practical Takeaway

We’re entering an era of multimodal foundation models, not just bigger language models. For text-heavy tasks, LLMs remain fantastic. For raw sensor streams from wearables or large structured datasets, dedicated models are already superior and far more practical.

Google’s SensorFM and TabFM show that the “foundation model” approach works brilliantly outside pure language. They deliver strong performance with minimal (or zero) task-specific training and handle real-world messiness gracefully.

If you work with health sensor data or tabular analytics, these models are worth watching closely — and using. The era of forcing everything through a text-only lens is starting to give way to more appropriate tools for the job.

---

Also read:

• 2026 SDG Global Business Forum Focuses on Business and Sustainable Development
• U.S. Small Business Employment Dips Slightly in June 2026 per QuickBooks Index
• IBM Shares Plunge 25% on Q2 Earnings Warning Amid AI Capex Shifts
• Control the Ideas, Not the Code
• Anthropic Explained When to Tinker with Effort Level in Claude Code (Spoiler: Not Often)

---

Thank you!