MiniMax Launches M3.1 Flash: A 428B MoE Agentic Model at $0.15 per 1M Input Tokens
MiniMax has released MiniMax M3.1 Flash on its API, priced at $0.15 per 1M input tokens. The company also confirmed that Space Bunny, a stealth model that processed over 200 trillion tokens, was an earlier M3.1 Flash build. For developers, the pitch is straightforward: near-frontier agentic performance at Flash-tier cost. Model weights are scheduled for release in late October 2026.
TL;DR
- Size: 428B total parameters with roughly 23B active (mixture-of-experts); 1,000,000-token context window
- Availability: MiniMax API, M Plan, and MiniMax Code today. Self-hosting hardware requirements have not been disclosed; weights arrive October 23, 2026
- Performance: Near-frontier on agentic office work, mid-pack on hard coding tasks — all figures are MiniMax-reported
- Best result: Automation-Bench v1.0.6 at 62.2, the top score in MiniMax's comparison table
- Worst result: Terminal Bench 4.0 at 38.1, versus 51.8 for Opus 5
What Is MiniMax M3.1 Flash?
MiniMax M3.1 Flash is a multimodal mixture-of-experts (MoE) model from MiniMax, targeting coding, agentic office work, and visual reasoning. MiniMax's product page lists 428B total parameters with approximately 23B active per forward pass, alongside a 1,000,000-token context window. The sparse MoE architecture keeps inference costs low relative to the model's total size, which is central to the $0.15 per 1M input token price point.
Positioned for the Agentic Era
The launch lands as agentic workflows — tool use, multi-step task execution, and computer interaction — become the primary benchmark battleground for frontier labs in 2026. MiniMax is positioning M3.1 Flash as a cost-efficient option for developers building automation pipelines and coding agents, where high-volume token consumption makes per-token pricing a decisive factor.
Availability and Next Steps
Developers can access M3.1 Flash today through the MiniMax API, M Plan, and MiniMax Code. Open weights are expected on October 23, 2026, though self-hosting hardware requirements remain undisclosed. Teams evaluating the model should weigh its strong agentic office-work scores against weaker hard-coding results before committing it to production pipelines.
via MarkTechPost