MiniMax M3: The Open-Weight Model That Runs Million-Token Contexts
Agentic AIAn open-weight model that sustains a one-million-token context while matching near-frontier performance on software engineering benchmarks is not a hypothetical capability. MiniMax M3 is it — and the weights are heading to Hugging Face within ten days.
What Happened
MiniMax released M3, an open-weight multimodal model with a 1 million token native context window — the first open-weight model to hit this threshold at proprietary-class benchmark performance. According to The Decoder’s coverage, M3 scores 59% on SWE-Bench Pro, sitting just behind Anthropic’s Claude Opus 4.7 (61%) but ahead of Gemini 3.1 Pro. On autonomous web search tasks (BrowseComp), M3 pulls ahead of Opus 4.7 at 83.5 versus 79.3.
The technical basis for the context performance is a new attention mechanism called MiniMax Sparse Attention (MSA), which filters relevant token blocks before running full attention calculations, reducing compute quadratically. MiniMax claims the result is four times faster performance than competing implementations at equivalent context lengths. The model handles 12–24 hour autonomous task execution, including self-correction and debugging across extended timeframes.
API pricing scales with context depth: requests under 512K tokens at standard rate, above 512K at a higher tier, with monthly plans from $20 (1.7B tokens) to $120 (9.8B tokens).
Why It Matters
The significance of M3 is less about beating any single benchmark and more about what the combination of open weights, million-token context, and near-frontier performance means for the agentic AI stack. Until recently, tasks requiring full codebase context, multi-session memory, or long document analysis across an entire enterprise corpus required a proprietary frontier model with significant API cost. M3 breaks that dependency.
For teams building agentic systems, the ability to run a million-token context on an open-weight model changes the cost and control calculus of long-horizon tasks. Code review across large monorepos, document analysis over full regulatory filings, multi-turn agents that maintain rich context across sessions — these use cases have been held back by either the context limitations of open models or the API costs of proprietary ones. M3’s sparse attention architecture addresses both constraints simultaneously.
The 12–24 hour task execution capability also signals that MiniMax is targeting the same autonomous agent workflows that Google’s SRE agentic AI deployment and other enterprise implementations are building toward — extended, largely unsupervised operation with self-correction rather than single-shot responses.
The So What
Open-weight models have historically required meaningful trade-offs against proprietary performance at the frontier. M3’s benchmark position and context length suggest the gap is narrowing faster than most enterprise AI procurement timelines assumed. For teams that have delayed building on open-weight infrastructure because of context limitations or benchmark gaps, M3’s release is a concrete reason to revisit that decision. The weights will be on Hugging Face shortly — the evaluation window is now.
Source: The Decoder
Content created with AI assistance and reviewed for accuracy.
Join the conversation
Stack Insiders is our free community for readers who want to go deeper — share resources, ask questions, and connect with others across every vertical we cover.
Join Stack Insiders →