Inference-Time Compute Hackathon is a 24-hour, in-person hackathon in San Francisco focused on pushing the frontier of what AI systems can do at inference time.

Hosted by Anthropic, Etched, Cognition, and Mercor, with compute support from Prime Intellect, this hackathon brings together ambitious builders to create projects that unlock new capabilities with modern models.

Teams will compete for a $50k top prize and $100k+ in total prizes. Participants will also receive serious compute support, including 8× H100s per team, meals, snacks, workspace for the full 24 hours, and $100 in compute credits from Prime Intellect.

This is an in-person hackathon, and participants are expected to be present in San Francisco for the full event.

Build multi-step AI systems that can take a goal and execute. Projects may include tool use, planning, long-horizon workflows, autonomous research, coding agents, or other agentic systems.

Build systems that evaluate people, models, or performance. Projects may include autograders, preference-ranking systems, rubrics, verified-skill signals, human-in-the-loop evaluation workflows, or applied AI tools for marketplaces and hiring.

Optimize open source models to challenge limits or imagine near-infinite compute, near-zero latency. Build as if that world is today. 

Apply solo or as a pre-formed team of 2–4. Solo applicants may be matched with teammates before the event.

Come ready to build something incredible. You’ll have 24 hours, high-end compute, food, workspace, and access to judges and founders from Etched, Cognition, Prime Intellect, Mercor, and other leading AI companies.

Join us in San Francisco to build the next frontier of inference-time AI.

• Vision & Problem Fit:
  Why is this problem worth solving, and does the approach actually fit the problem?
• Technical Execution:
  How well is the solution built? We care about what works, not just what's described
• Novelty & Insight:

  What are the new insights this solution is bringing to the table, anything outside of a rehash of something that already exists?
• Impact & Trajectory:
  If this worked at full scale, how much would it matter?
• Presentation & Demo:
  Can you clearly explain what you built? Does it work / to what extent does it work?