From Mixture of Experts to Mixture of Agents with Super Fast Inference
summary
Hands-on workshop shows how to assemble a Mixture of Agents (MoA) system using Qwen3-32B and Llama 3.3-70B running on Cerebras hardware, claiming MoA can beat frontier models on some benchmarks by routing tasks across specialized agents. Demonstration includes a Streamlit app and an automated grader for a Python optimization challenge that scores up to 120, a 7.4s solution to a math problem versus 293s for a reasoning model, and hardware performance of 15.5x faster for Llama 3.3-70B inference.
AI-written summary · about
chapters (21)
- 0:00Introduction and API setup
- 0:36Workshop agenda and goals
- 1:43Speaker introductions
- 3:48Cerebras hardware advantages
- 5:06Mixture of experts concepts
- 7:21Implementing mixture of experts
- 10:31Introduction to mixture of agents
- 12:29Applications of mixture of agents
- 14:27How mixture of agents works
- 16:05Cerebras hardware architecture
- 18:42Specializing agents for complex tasks
- 20:15Workshop hands-on setup
- 22:41Workshop app walkthrough
- 25:44Initial workshop experimentation
- 29:14The competition challenge
- 36:23Competition results and prizes
- 37:32Q&A session: Scaling and roadmap
- 40:50Q&A session: Efficiency and engineering
- 43:44Q&A session: Custom models and fine-tuning
- 45:25Q&A session: Future developments
- 47:05Workshop conclusion and final questions
related talks
46mArchitecting Efficient Agents with Semantic Routing and Mixture-of-Tools
47mOwn Your LLM: Running Generative and Agentic AI Locally
27mAI Demystified Unveiling the Intelligence Behind the Machine
37mOne agent, one job, better A
40mBuilding Agents with Agent Bricks and MCP
40mAgents in Action: Advancing Open Source AI Missions
41mThree agents, three frameworks, one talk
16mUnleash the superpowers of the AI PC