March 2026 Summaries

Runway has partnered with Modal to provide real-time inference infrastructure for Runway Characters, an API that enables developers, businesses, and consumers to create customizable conversational video agents with controllable appearance, voice, personality, knowledge, and actions. Built on Runway’s GWM-1 world model, Characters can generate expressive digital personas from a single image without fine-tuning and is already used by organizations across technology, entertainment, advertising, and gaming for applications such as customer support, training, advertising, and immersive virtual worlds. Modal’s serverless GPU platform supports the low-latency, globally distributed, and variable-demand requirements of sustained video conversations, allowing Runway to progress from a proof of concept to production in under 30 days. Its multi-node GPU clusters, RDMA networking, regional routing, and automatic scaling help maintain real-time expressions, lip-sync, and gestures for users worldwide without Runway directly managing regional infrastructure. Runway Characters is now available to developers and businesses through its developer platform and to consumers through Runway’s website.

Doppel, an AI-native cybersecurity platform focused on detecting and disrupting social engineering attacks, migrated much of its machine learning workflow to Modal to accelerate experimentation and simplify real-time model deployment. Previously, sequential training runs, lengthy jobs, costly failures, and bundled changes slowed experimentation, while Modal enabled parallel execution of independent tasks such as cross-validation through standard Python constructs, improving the speed of evaluating hypotheses. The company also uses coding agents for operational tasks including launching experiments, gathering metrics, and summarizing outcomes, while machine learning engineers retain responsibility for selecting worthwhile ideas. For inference, Doppel moved away from a GCP Cloud Run workflow involving custom Docker containers, single-GPU instances, Flask endpoints, and potentially slow builds or scaling delays. Modal’s image caching, persistent model-weight volumes, serverless scaling, and direct function invocation reduced deployment times, integration code, and infrastructure management, allowing the team to focus more on developing and assessing detection models.

Modal announced Directory Snapshots, which allow users to capture and reuse individual directories across Sandboxes independently of base images, supporting faster initialization and separation of dependencies from application code. The company is also offering free access through the end of April to Z.ai’s GLM-5 open-weights model, designed for long-horizon and coding agents and compatible with several popular agent tools. Billing improvements include a refreshed interface and generally available CLI and API reporting tools for Team and Enterprise workspaces, while SDK updates add changelog queries, dashboard shortcuts, and Sandbox cleanup controls. Modal additionally highlights a webinar on scaling coding-agent infrastructure, research projects using its GPU and Sandbox environments to automate AI development, and several upcoming virtual and in-person events focused on AI, agents, and infrastructure.