Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It
Conference Context
- Date/time: 2026-06-30 · 2:50pm-3:10pm
- Track/room: Posttraining & Midtraining · Track 9
- Speaker(s): Olive Song, Dan Fu
- Session type/status: session · confirmed
- Track: Posttraining & Midtraining
- Room: Track 9
- Session type: session
- Status: confirmed
Session Description
Olive Song (RL Lead, https://www.minimax.io/) and Dan Fu (VP of Kernels, https://www.together.ai/) dig into the engineering behind one of the most widely used open model families in the agent ecosystem: how MiniMax built the model for agentic workloads, and what it takes to serve it at scale. Olive on the model side: The RL decisions behind long-context reasoning and tool use What training for agentic behavior actually looks like in practice Dan on the infrastructure side: Why agentic workloads break inference engines built for chat: prefill-heavy traffic, high cache hit rates, long-context inputs The kernel-level optimizations built for MiniMax's workload profile How the two teams collaborate on model launches and ongoing performance work
Media Evidence
Minimax M2: Building the #1 Open Model – Olive Song, MiniMax (speaker-match related prior/adjacent AI Engineer video; captions: English auto-captions).
- Source video:
youtube-lY1iFbDPRlw - Slide deck: Dense Slides: Minimax M2: Building the #1 Open Model – Olive Song, MiniMax — 5 visible slide image(s); 5 HTML recreation(s).
- Additional slide evidence: Slides: Minimax M2: Building the #1 Open Model – Olive Song, MiniMax, Reconstructed Slides: Minimax M2: Building the #1 Open Model – Olive Song, MiniMax
- Slide-derived themes for
youtube-lY1iFbDPRlw: excel, speech, music, users, globally, olive, song, senior.

Evidence Graph
This evidence graph is generated from currently linked source material: official schedule text, related video pages, cached transcripts, visible slide text, dense/reconstructed slide pages, and AI slide-classification audits.
Media Signals
youtube-lY1iFbDPRlw— 7 slide-derived text signals- Slide-derived themes for
youtube-lY1iFbDPRlw: excel, speech, music, users, globally, olive, song, senior. - Evidence links for
youtube-lY1iFbDPRlw: youtube lY1iFbDPRlw, youtube lY1iFbDPRlw slides, youtube lY1iFbDPRlw dense slides, youtube lY1iFbDPRlw reconstructed slides
Agent Reading Notes
Use these signals to refine the synopsis, topic links, people/company context, and method notes. If a source is a related external video rather than an exact official recording, keep it framed as supporting evidence.
Transcript Status
Related video transcript availability: English auto-captions. Treat this as supporting context, not a recording of this exact scheduled session unless later confirmed. Not fetched yet.
People
Supporting Slides
- youtube lY1iFbDPRlw slides — extracted from the related public AI Engineer video.
Slide Evidence
- Slide-only cropped deck: youtube lY1iFbDPRlw dense slides (5 viable slide images).
- Related slide/OCR pages:
- youtube lY1iFbDPRlw dense slides
- youtube lY1iFbDPRlw reconstructed slides
- youtube lY1iFbDPRlw slides
- Slide-derived terms:
minimax,research,open-weight,senior,engineer,tasks,minimax-m2,agentic,model,fast,coding,intetgence,everyene,olive,song,intelligence,real,experience
Synthesis
Synthesized Breakdown
Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It ## Conference Context - Date/time: 2026-06-30 · 2:50pm-3:10pm - Track/room: Posttraining & Midtraining · Track 9 - Speaker(s): Olive Song, Dan Fu - Session type/status: session · confirmed - Track: Posttraining & Midtraining - Room: Track 9 - Session type: session - Status: confirmed ## Session Description Olive Song (RL Lead, https://www.minimax.io/) and Dan Fu (VP of Kernels, https://www.together.ai/) dig into the engineering behind one of the most widely used open model families in the agent ecosystem: how MiniMax built the model for agentic workloads, and what it takes to serve it at scale. Olive on the model side: The RL decisions behind long-context reasoning and tool use What training for agentic behavior actually looks like in practice Dan on the infrastructure side: Why agentic workloads break inference engines built for chat: prefill-heavy traffic, high cache hit rates, long-context inputs The kernel-level optimizations built for MiniMax's workload profile How the two teams collaborate on model launches and ongoing performance work ## Media Evidence Minimax M2: Building the #1 Open Model – Olive Song, MiniMax (speaker-match related prior/adjacent AI Engineer video; captions: English auto-captions). - Source video: youtube-lY1iFbDPRlw - Slide deck: Dense Slides: Minimax M2: Building the #1 Open Model – Olive Song, MiniMax — 5 visible slide image(s); 5 HTML recreation(s).
- Additional slide evidence: Slides: Minimax M2: Building the #1 Open Model – Olive Song, MiniMax, Reconstructed Slides: Minimax M2: Building the #1 Open Model – Olive Song, MiniMax - Slide-derived themes for youtube-lY1iFbDPRlw: excel, speech, music, users, globally, olive, song, senior.
Speaker And Company Context
- Olive Song — RL Lead at MiniMax.
- Dan Fu — VP of Kernels at Together AI.
Topics Covered
Derived Links And Source Material
- youtube lY1iFbDPRlw — related YouTube source page.
- youtube lY1iFbDPRlw slides — slide evidence.
- youtube lY1iFbDPRlw reconstructed slides — slide evidence.
- youtube lY1iFbDPRlw dense slides — slide evidence.
Novel Concepts / Clever Methods
- No highlighted novel concept has been detected yet.
Evidence Boundary
This synthesis is based on the official schedule and linked source pages. It should be revisited when exact session recordings or transcript-backed secondary sources are available.