Sign InOpen Brain
AI EngineerVideoSource Linked

Why AI Agents Need Million-Token Context — Thomas Wolf & Olive Song, MiniMax

MiniMax argues that long-running, multimodal agents need enough context for repeated tool responses, while sparse attention makes that context more tractable.

AI Engineer · Sep 4, 2026
Open Source Open MarkdownOpen JSON
Source Summary

MiniMax M3 has roughly **400B total parameters**, activates **20B**, and supports text, images, and video. Its **1M-token context** uses MiniMax Sparse Attention: an index branch selects relevant blocks, then a sparse branch performs attention over them.

Practical Implication

For agent builders, the practical claim is that multi-round tool responses and long, unstructured inputs can exceed short contexts. M3’s native multimodal training also points toward agents that inspect reports, presentations, tutorials, images, and video before acting with tools.

Agent-Ready Context
MiniMax M3 has roughly **400B total parameters**, activates **20B**, and supports text, images, and video. Its **1M-token context** uses MiniMax Sparse Attention: an index branch selects relevant blocks, then a sparse branch performs attention over them.

For agent builders, the practical claim is that multi-round tool responses and long, unstructured inputs can exceed short contexts. M3’s native multimodal training also points toward agents that inspect reports, presentations, tutorials, images, and video before acting with tools.

A large context window does not establish useful recall across its full length, and the discussion supplies no retrieval or agent-task measurements. Trillion-token attention is presented only as a research direction requiring architecture and hardware work.
Connected Context · Feed7 Judgment

This signal supplies an architectural rationale for million-token, multimodal agents: accumulated tool output and unstructured media may require sparse attention rather than simply larger dense windows. It does not validate full-window recall or agent performance, so it reinforces—not resolves—the candidates’ recurring requirement for matched workload tests before treating context capacity as useful capability.

Gemini 3.8 Flash now available on AI GatewayGemini provides another million-token multimodal agent route, but neither signal demonstrates that nominal capacity translates into reliable recall or better task completion.Hy4 Preview now available on AI GatewayHy4 confirms that open MoE models are also entering the million-token coding-agent category, making comparative sustained-context testing more important than window-size comparison.GLM 5.3 now available on AI GatewayGLM 5.3’s long-horizon engineering claims are exactly the kind of agent-level evidence needed, but its supplied record still lacks comparative results that would validate the broader value of a 1M-token window.What's Next After RLHF? — Diogo Almeida, TypeSafe AIThe RLHF critique adds a reliability boundary: expanded context and multimodal input increase available evidence, but do not by themselves make autonomous decisions calibrated or dependable.
Context Map
modelcodingimage#open-models#reasoning#coding-agents
Uncertainty
A large context window does not establish useful recall across its full length, and the discussion supplies no retrieval or agent-task measurements. Trillion-token attention is presented only as a research direction requiring architecture and hardware work.