Best use of Modal

Take your Junction project from laptop to production AI on Modal.

Description

Every Junction team is building with AI, and most hit the same wall: the model runs on a laptop, but too slowly, not on a GPU, and not behind a URL the judges can open. Modal gets you past that wall. Use Modal to power the project you're already building for your main Junction challenge: serve an open model on a GPU, run your agent's code in secure Sandboxes, fine-tune a model on your own data, or fan out a big batch job, all in plain Python with no Dockerfiles, Kubernetes or YAML. We're looking for the most ambitious working AI system built on Modal, where Modal does the heavy lifting and the team spends its 48 hours on the problem. Inference, agents, sandboxes and training: your whole AI stack on one platform.

Mentors:

Engineers and Developer Relations from Modal with experience in inference, AI agents and code-execution sandboxes, fine-tuning and reinforcement learning, and real-time audio and video pipelines. We'll be at the Modal booth throughout the event and on the Junction Discord for debugging help.

Main contact person: Filip Makraduli, Developer Relations Engineer

Modal was built so that a single developer can do what used to take a platform team, which makes it a natural fit for a 48-hour hackathon. Best use of Modal is a tech challenge, so you compete in it on top of your main challenge: use Modal to power the project you're already building and you can enter. The same project can enter other tech challenges too.

What we’ll bring:

Every participant gets $100 in Modal credits for CPUs and GPUs, up to the latest NVIDIA B300s, billed per second and scaling to zero when idle. Modal is Python-native: decorate a function, run modal run, and it executes in the cloud in seconds; modal deploy turns it into an autoscaling web endpoint with a public URL ready for your demo. Teams can serve open-weight LLMs, vision-language and speech models; run untrusted, LLM-generated code in isolated Sandboxes; fine-tune or RL post-train models on one GPU or a multi-node cluster with Modal Dojo, our open-source training library; and store weights and datasets in Volumes. We'll provide a Junction starter kit of ready-to-run examples. The kick-off workshop will be at the Modal booth and announced on Discord, and Modal engineers will be on site all weekend.

Access and list of technologies we are bringing:

• Modal compute credits ($100 per participant) for CPUs and GPUs: T4, L4, A10, L40S, A100 (40 and 80 GB), RTX PRO 6000, H100, H200, B200 and B300 • Modal Python SDK and CLI (pip install modal), plus JavaScript and Go SDKs for Sandboxes • Modal Sandboxes: secure, on-demand containers for agents, code interpreters and evals • Web endpoints (FastAPI, WebSockets, WebRTC) with autoscaling and public URLs • Volumes, Dicts and Queues for model weights, datasets and shared state • Modal Dojo: open-source library for fine-tuning and RL post-training open models on Modal GPUs • Starter examples: LLM serving with vLLM and SGLang, speech-to-text, text-to-speech, vision, RAG, agents and RL • On-site mentors from Modal engineering and Developer Relations

To claim $100 in Modal credits you need to sign up for a Modal account, then redeem the credits via the link (https://modal.fillout.com/t/qMXCmRGseUus?code=VVX-2QX-8HW) using the code VVX-2QX-8HW.

The credits are valid for one year from redemption, so everyone can keep building well after the hackathon wraps up.