MiniMax plans 2.7 trillion-parameter open-weight AI model, the largest from a Chinese firm

Chinese AI startup MiniMax Group is developing a large language model with 2.7 trillion parameters that it plans to release as open-weight software as early as this quarter, according to a report by Reuters on Tuesday, in what would be the largest open-weight model ever released by a Chinese company and potentially the largest in the world.reuters

A New Frontier in Open AI

The model, internally codenamed M3 Pro, dwarfs MiniMax's current flagship M3 model, which has 428 billion parameters — making the new system roughly six times larger. A person with direct knowledge of the plans told Reuters the release could come in the third quarter of 2026, though the company declined to comment. The Information first reported on the development on July 8.kucoin+3

If released as planned, M3 Pro would surpass existing trillion-parameter open-weight models, including DeepSeek-V4-Pro at 1.6 trillion parameters. MiniMax already operates one of the most capable open models available: MiniMax-M1, which it describes as the world's first open-source large-scale hybrid-attention reasoning model, built on 456 billion total parameters with a one-million-token context window and released under the Apache 2.0 license.gigazine+2

Chinese Labs Accelerate Open Releases

The announcement arrives amid a broader push by Chinese AI laboratories to release powerful models under permissive licenses. Tencent officially launched Hy3 on July 6, a 295-billion-parameter mixture-of-experts model with 21 billion active parameters, released under Apache 2.0. Tencent said Hy3 delivers performance comparable to flagship models with two to five times its parameter count, recording scores rivaling GLM-5.2 and DeepSeek-V4-Pro on reasoning benchmarks while outperforming GPT-5.5 on scientific research tasks.tencent+1

The wave of open releases from Chinese firms has intensified competition with Western AI labs, many of which have moved toward closed or restricted licensing for their most capable models. MiniMax, which is publicly listed in Hong Kong, also plans to launch H3, its frontier-level multimodal video generation model, later this month, according to the same source cited by Reuters.reuters

What Comes Next

The scale of M3 Pro raises questions about the computational resources required to run such a model and how broadly it can be deployed outside major research labs. MiniMax's track record with M1 — which achieved strong benchmark results at a reported training cost of just $534,700 — suggests the company will emphasize inference efficiency alongside raw parameter count. The final release name for the 2.7 trillion-parameter model has not yet been determined.venturebeat+1