PIF-Backed HUMAIN Launches Humain-M3 Arabic Model at LEAP Riyadh

HUMAIN, the Public Investment Fund-backed artificial intelligence company, announced humain-m3 on September 3, 2026, unveiling the frontier Arabic-language model at LEAP in Riyadh and releasing it in research preview through HUMAIN Node. Commissioned by HUMAIN and delivered by MiniMax, the model was further pre-trained on more than one trillion tokens of Arabic-native content.

humain-m3 is a 428-billion-parameter mixture-of-experts model built on the MiniMax-M3 lineage. According to HUMAIN’s announcement, the model achieved the highest average score among the frontier models tested across seven public Arabic benchmarks, with strong performance across Arabic language understanding and reasoning.

Arabic Benchmark Results

HUMAIN’s evaluation put the previewed humain-m3 checkpoint at an average of 89.37% across the seven benchmarks, equally weighted, ahead of GPT-5.6 SOL at 87.30%, Opus 5 at 87.34%, and the MiniMax M3 reference checkpoint at 80.34%, according to the results table published on the HUMAIN Node model page. HUMAIN said the previewed model leads five of the seven benchmarks against the strongest frontier models it evaluated.

The suite spans core understanding, native and translated knowledge, academic examinations, language proficiency, truthfulness, and retrieval-augmented generation. In HUMAIN’s results, humain-m3 scored 86.45% on AlGhafa for core Arabic understanding, 90.70% on ArabicMMLU for native-Arabic broad knowledge, 67.67% on Arabic EXAMS for academic examinations, 95.44% on MadinahQA for Arabic language proficiency, 97.53% on AraTrust for truthfulness and trust, 94.63% on ALRAGE for retrieval-augmented generation, and 93.20% on Translated MMLU. The company said its Arabic post-training adds nine points on average over the reference model the system builds on, and described the scores as its own evaluation of the previewed checkpoint.

Architecture and Training

The Node page describes humain-m3 as activating 23 billion parameters per token and characterizes the underlying architecture as a natively multimodal mixture-of-experts engineered for agents. HUMAIN said the model handles frontier tool use and computer use for long-horizon agent workflows in Arabic and English, offers three thinking modes — always-on, adaptive, and off — and was trained jointly on text, image, and video from the first step, with long-video understanding and native screen operation.

Access Through HUMAIN Node

The research and evaluation preview is available through HUMAIN Node, the company’s platform giving developers, researchers, and enterprises access to advanced AI models and inference capabilities. Users can try the model in a no-code playground or call humain-m3 through an OpenAI-compatible API endpoint.

HUMAIN structured access in tiers. Creating an account opens access requests for the model. A limited-preview tier provides the model with a Saudi alignment guardrail applied, with thinking and streaming off and some added latency. A research-preview tier provides the full checkpoint with thinking, streaming, and lower latency. HUMAIN said the preview period lets it evaluate capabilities, safety, and alignment across Arabic dialects with early users before general availability, and that user feedback shapes the next checkpoints.

The model is the first open to preview users on HUMAIN Node, where the company said more than 100 frontier, proprietary, and open-source models from leading providers are already on the platform with broader access opening soon. Node offers a single API, key, and bill across models, spending limits by team and project, and a choice of global, in-Kingdom, or sovereign hosting.

Planned Open-Weight Release

HUMAIN said it expects the model weights to be released under the MiniMax Community License once safety training and alignment are complete, a step it currently targets for next month. The company framed the preview as a way for researchers and developers to test the model and provide feedback ahead of that release.

“Arabic is spoken by hundreds of millions of people, yet it remains significantly underrepresented at the frontier of artificial intelligence,” said Tareq Amin, HUMAIN’s CEO. “With humain-m3, we are investing in changing that.”

HUMAIN said the launch combines two elements of its intelligence strategy: developing world-class AI capabilities for Arabic and making frontier intelligence more accessible through HUMAIN Node. Alongside its own models and products, including the ALLAM family of Arabic models, the company said it is investing in and working with leading AI companies globally to expand access to advanced AI capabilities.