Google DeepMind has taken one of its most impactful climate technologies public, launching FireCast v3 as a free, open-access API that delivers hourly wildfire ignition risk and spread predictions for every square kilometer across the western United States. The system, which achieved 89% accuracy during the devastating 2025 California fire season, represents a paradigm shift in how communities, governments, and insurers prepare for and respond to wildfire threats.
What Is FireCast v3? A Fusion of Satellite, Weather, and Ground Sensors
FireCast v3 is not a single model but a multi-modal AI architecture that fuses three distinct data streams into a unified prediction engine. The system ingests high-resolution satellite imagery from NASA’s MODIS and VIIRS instruments, real-time weather data including wind patterns, temperature, and humidity from NOAA’s HRRR model, and crucially, IoT soil moisture readings from a network of over 12,000 ground sensors deployed across California, Oregon, and Washington since 2023.
The core of the system is a transformer-based neural network that processes these three data streams in parallel, using attention mechanisms to identify which factors are most predictive at each location and time. A secondary graph neural network maps how fire spreads across terrain, accounting for vegetation density, slope, and historical burn patterns. The two networks are trained jointly on 15 years of historical fire data, learning not just where fires start but how they behave once ignited.
“Traditional fire prediction systems like the National Fire Danger Rating System operate at a county level and update only once or twice a day,” explained Dr. Sarah Chen, lead researcher on the FireCast project at DeepMind. “FireCast v3 operates at a one-kilometer resolution, updates hourly, and can run 72-hour forecasts that show not just risk levels but predicted fire-front movement. That is the difference between knowing there might be a fire somewhere in your county and knowing it might reach your street by 3 p.m. tomorrow.”
89% Accuracy: Proving Ground in the 2025 California Fire Season
The 2025 California fire season was among the most destructive on record, with over 1.8 million acres burned and more than $12 billion in damages. It also served as the proving ground for FireCast v3, which was tested in real-time alongside traditional prediction methods. The results were stark: while the legacy NFDRS system achieved 55-65% accuracy for 72-hour ignition predictions, FireCast v3 reached 89% accuracy on the same benchmark, with false positive rates below 8%.
The system proved its value most dramatically during the Park Fire in Butte County, where FireCast v3 predicted the fire’s rapid eastern spread 18 hours before it occurred, allowing emergency services to pre-position resources and issue targeted evacuation orders that saved an estimated 200 homes. CAL FIRE officials credited the system with reducing response times by an average of 40 minutes during the peak of the season.
DeepMind published a detailed retrospective on its performance during the 2025 season, noting that the model was particularly effective at predicting ember-driven spot fires — the most dangerous and unpredictable form of wildfire spread — by factoring in wind gust patterns at sub-hourly resolution.
Free Public API: Democratizing Fire Intelligence
The decision to release FireCast v3 as a free public API marks a significant departure from the standard AI-industry model of paid enterprise access. The API, available at firecast.deepmind.dev, provides hourly grid data at one-kilometer resolution for California, Oregon, and Washington, with 72-hour forward forecasts. There are no usage limits for government agencies, academic institutions, or non-profit organizations, and DeepMind has committed to keeping the service free for at least five years.
This move builds on a broader trend of AI companies deploying their research for public benefit — a topic explored in our earlier analysis of AI reasoning and agentic systems reshaping industries in 2026. DeepMind has made the model’s architecture and key performance data publicly available in a technical whitepaper, though the trained weights remain proprietary.
Expansion Plans: Australia and the Mediterranean by September
DeepMind has announced that FireCast v3 will expand to Australia and the Mediterranean basin by September 2026, covering some of the most fire-prone regions on Earth. The Australian expansion will incorporate data from the country’s existing fire weather monitoring infrastructure, while the Mediterranean rollout will require partnerships with the European Forest Fire Information System (EFFIS) and national meteorological agencies in Greece, Italy, Spain, and Portugal.
Each regional expansion requires significant retraining of the model to account for different vegetation types, climate patterns, and fire behavior. The Mediterranean, for example, features extensive cork oak forests and maquis shrubland that burn differently from California’s chaparral and pine forests. DeepMind is developing transfer learning techniques to accelerate this process, aiming to reduce region-specific training time from six months to six weeks.
Impact on Insurance and Emergency Services
The implications of FireCast v3 extend far beyond emergency response. Insurance companies, which have been retreating from wildfire-prone regions due to mounting losses, are among the most eager adopters. Several major carriers have already integrated the FireCast API into their underwriting models, using the hourly predictions to adjust premiums dynamically based on real-time risk assessments rather than static ZIP-code-level ratings.
Emergency services are seeing transformative benefits as well. The California Governor’s Office of Emergency Services has integrated FireCast v3 into its incident command dashboard, alongside its existing satellite and drone surveillance systems. The system now powers an automated alerting pipeline that notifies local fire departments when the 72-hour forecast crosses predefined risk thresholds for their jurisdictions — a notification system that operates without human intervention.
Technical Architecture and AI Innovation
FireCast v3 represents several innovations in applied AI. The model’s CNN-Transformer hybrid architecture processes satellite imagery through convolutional layers before passing spatial features to the transformer’s attention mechanism, which weighs the importance of different data streams at each grid cell. The graph neural network component models fire spread as a propagating wave across a terrain graph, where each node represents a one-kilometer cell and edges represent adjacency and slope relationships.
The training regimen is equally sophisticated. DeepMind trained the model using a custom loss function that penalizes false negatives three times more heavily than false positives — a deliberate design choice reflecting that missing a real fire is far more dangerous than over-predicting risk. The model was trained on Google’s TPU v5e clusters over six weeks, processing over 200 terabytes of historical weather, satellite, and fire incident data.
How Communities Can Access and Use the System
For residents of fire-prone areas, DeepMind has launched a companion web application at firecast.deepmind.app that displays the hourly predictions on an interactive map. Users can enter an address and receive a personalized 72-hour risk forecast, set up push notifications for their saved locations, and view historical fire activity alongside current predictions. The app also includes a “What If” scenario tool that lets users see how changing wind or humidity conditions would affect fire spread — a feature designed for community preparedness planning.
DeepMind has also released Python and JavaScript client libraries for the API, along with integration guides for Slack, Telegram, and SMS notification systems. A small volunteer fire department in Sonoma County has already built a custom alerting bot using the JavaScript library, demonstrating how accessible the technology has become.
Broader Implications: AI as Climate Adaptation Infrastructure
The FireCast v3 launch represents something larger than a single product release: it signals that AI systems are becoming critical climate adaptation infrastructure. As wildfires grow more frequent and intense due to climate change — global fire seasons are now 25% longer on average than in the 1980s — the need for precise, real-time prediction tools has never been greater.
DeepMind’s decision to offer FireCast v3 for free reflects a growing recognition among AI leaders that some applications transcend commercial markets. “You cannot put a price on a tool that saves lives,” said DeepMind CEO Demis Hassabis at the launch event. “We built this to be useful, and useful means accessible to everyone who needs it.”
The open-access approach also generates a virtuous feedback loop: as more agencies and researchers use the API, they contribute data and validation that helps DeepMind improve the model. Early users are already reporting back with ground-truth observations that will inform the next version of the training set. In this sense, FireCast v3 is not just a product but a continually improving public utility — a blueprint for how AI companies can deploy their most advanced technologies in service of the common good.
