MiniMax preps 2.7T-parameter M3 Pro for Q3 open-source drop

The most interesting story in Chinese AI this week is not about a model that shipped, but about one that hasn't. The Information reported, via a Techmeme brief, that MiniMax is working on a 2.7T-parameter model internally called M3 Pro, which it could release as early as Q3 as open source. If those numbers hold, it would be larger than any other Chinese AI model currently on the market.

The scale delta is what makes people look twice. MiniMax's current M3, released on June 1, 2026, has a total parameter count of up to 428B, with 23B active parameters at inference. The reported M3 Pro would be over 6x bigger than that on total parameters. The company spent this cycle positioning M3 as a low-cost frontier alternative, pricing it at $0.12 per 1M input tokens compared with $5 for Opus 4.7, and pairing it with a one-million-token context window and native multimodality. Going to 2.7T is a different bet, planned within roughly a quarter of M3's June debut.

Why this matters if you don't work at a lab: a 2.7T open release, if it lands as described, resets what "available to download and run yourself" looks like at the top end, and puts real pressure on any government or enterprise that assumed frontier scale would stay gated by API.

The honest caveat is that this is single-sourced reporting attributed by Juro Osawa at The Information to two people with knowledge of the plan. Parameter counts are not the same as capability, sparse or MoE architectures could make the 2.7T figure less dramatic at inference than it sounds on a headline, and the reporting does not say what active-parameter design M3 Pro uses, what license it will ship under, or what hardware you would actually need to serve it. Take the specifics as reported, not settled.

What to watch for is the license and the active-parameter split at release. Those two decisions will decide whether M3 Pro is a headline number or a genuinely usable open frontier for developers, researchers, and cost-sensitive buyers weighing open weights against paid APIs.