사우디 HUMAIN-M3은 미니맥스 M3 오픈 가중치 파운데이션 모델을 활용해 자국 주권 AI 스택을 구축하고 글로벌 베이스 모델 파트너십 경쟁에 나서고 있다.

At LEAP 2026 on September 3, Saudi Arabia’s PIF-backed AI company HUMAIN unveiled HUMAIN-M3, an Arabic-first frontier model currently available in preview on the HUMAIN Node platform, with open weights scheduled for release later. The model is built upon Shanghai-based startup MiniMax’s open mixture-of-experts (MoE) stack, which was launched approximately three months prior.

HUMAIN-M3 retains MiniMax M3’s MoE architecture—approximately 428 billion total parameters with roughly 23 billion activated per token—and continues training on more than one trillion native Arabic tokens. According to seven public Arabic benchmarks published by HUMAIN, the model secured five first-place finishes and achieved an equal-weighted average score of 89.37%. This represents a 9.03-point improvement over the base M3 model and surpasses cited performance metrics for both GPT-5.6 SOL and Claude Opus 5. The operational advantage centers on efficiency: leveraging the general reasoning, multimodal capabilities, agentic functions, and long-context handling of an open frontier foundation, then specializing it for local language and cultural requirements.

This methodology aligns with a broader global scramble for sovereign AI. Tracking cited in Chinese industry coverage indicates that by mid-2026, there were 184 government-backed sovereign AI initiatives across 67 countries, with 41 launched in the first half of the year alone. The deployment of open-weight models for localized continued training has more than doubled since late 2024. Procurement priorities now heavily emphasize dialect coverage, on-premises data control, and resilience against sudden export-control disruptions that could sever remote API access—factors that strongly favor downloadable weights paired with domestic post-training pipelines.

MiniMax’s international footprint is expanding along similar lines. The company reports its products now serve more than 230 countries and regions. In Europe, it has partnered with SambaNova to deliver high-speed inference on SN50 hardware and with Nebius Token Factory to provide a dedicated hosted open model. HUMAIN advances this ecosystem further by transitioning from simple weight downloads or API access into a comprehensive national model program, deploying proprietary language datasets and domestic compute to fine-tune the Chinese checkpoint.

For Chinese open-model laboratories, the primary customer is no longer merely a developer chasing leaderboard rankings. Governments increasingly seek foundational checkpoints they can own, continuously update, and operate independently. MiniMax M3’s integration into HUMAIN-M3 serves as an early, tangible example of this shift. It signals that reliable open-weight distribution, hardware optimization, and sustained technical support will carry as much strategic weight as incremental benchmark improvements. As competition expands beyond the U.S.-China duopoly, the defining metric is shifting toward which providers can deliver robust, reusable foundations rather than which models dominate English-language evaluations.