April 2025 Summaries

Modal has released alpha SDKs for JavaScript and Go, allowing server-side applications to call deployed Modal Functions and create secure, isolated sandboxes without using Python for client-side integration. The SDKs support workloads such as image processing, LLM inference, transcription, batch jobs, executing generated or untrusted code, running repository tests, and launching containers for one-off tasks, though deploying Modal Functions still requires Python. Users can authenticate through the Modal CLI or environment variables and install the packages through npm or Go modules, with examples demonstrating remote function calls and sandbox command execution. The initial release supports JSON-like and byte data types but lacks advanced capabilities including map and spawn modes, sandbox tunnels, volumes, and container lifecycle controls, which are planned for later. Modal describes the libraries as lightweight and consistent, notes successful testing at 10,000 concurrent sandboxes, and cautions that the alpha APIs and naming conventions may change as development continues.

Lemon Slice uses Modal to operate AI character video products, evolving from a viral tool that generated speaking-character videos from images and text or audio into Lemon Slice Live, which enables real-time video conversations with AI characters. Rather than managing custom AWS or GCP infrastructure, the company deployed its 1-billion-parameter video model through two Python-based Modal Functions, allowing it to scale to 10,000 requests per hour with autoscaling GPU containers, reduced initialization times, and parallel model evaluation. For its live product, Modal launches separate containers for a Pipecat real-time processing server and GPU video inference, while Pipecat coordinates speech recognition from Deepgram, conversational responses from Grok, speech synthesis from ElevenLabs, and video generation. Direct TCP communication, regional co-location, and Daily’s WebRTC streaming infrastructure help reduce latency, producing video-and-audio responses in approximately three to six seconds.

Sync, a research lab founded by the team behind Wav2Lip, develops foundational AI models for manipulating humans in video, including a zero-shot lip-syncing system that can preserve a speaker’s style across translated languages. After gaining attention with a viral Hindi-language video demo, the five-person team needed to move beyond Google Colab and infrastructure tools such as AWS Lambda, which created deployment, scaling, and GPU-support challenges, while Replicate’s container rebuild process slowed updates. Using Modal’s startup credits and developer workflow, Sync was able to deploy changes rapidly, relying on Modal for autoscaling and GPU infrastructure rather than MLops management. The company reports deploying up to 95 times daily, releasing 10 major production model variants and roughly 1,000 intermediate iterations in a year. Its production workflow processes more than 100 hours of video daily by splitting long videos into scenes, running parallel face detection and translation on T4 GPUs, applying its lip-sync model on A100 GPUs, and stitching the results together. Sync plans to expand beyond translation and lip-syncing into AI-powered emotion editing, pose adjustment, and changes to subjects’ physical characteristics.

Modal has expanded its asynchronous job capabilities by allowing up to one million inputs per Function, increasing `.spawn()` submission limits, and retaining FunctionCall results for seven days. Recent client improvements include support for ephemeral apps launched from containers, more reliable Dockerfile context handling, configurable image entrypoint arguments, and Git commit visibility in the CLI and dashboard, while Modal Client v1.0 is expected to introduce cleaner APIs and deprecation guidance. The company also released a TensorRT-LLM example for sub-400-millisecond language-model inference, video walkthroughs for deploying DeepSeek and OpenAI-compatible vLLM services, and highlighted customer launches from Imbue, Phonic, and Firebender using its infrastructure. Additional updates include recognition as the second-most-promising early-stage company on the 2025 Enterprise Tech 30 list, an open-source LLM demo event with Mistral, and a San Francisco billboard campaign.

Founded in 2021, Modal has updated its longstanding thin-lined cube “M” logo and wordmark to improve readability, scalability, and consistency across applications ranging from favicons to billboards while preserving the recognizable identity that inspired its mascots, Moe and Dal. Developed with designer Ty Wilkins, the revised branding coincides with Modal’s first out-of-home campaign in San Francisco, intended to expand the New York-based company’s presence in a major technology and AI hub. The campaign features three billboards across the city, each highlighting a different aspect of Modal’s AI infrastructure platform, and will remain displayed through the end of May.