Loading page…
The Saudi-backed company is offering developers early access now, while its strongest performance claims come from its own tests and its planned weight release remains contingent on safety and alignment work.
Listen to this story
HUMAIN is offering developers a research preview of humain-m3 through HUMAIN Node, an Arabic-specialized 428-billion-parameter mixture-of-experts model adapted from MiniMax-M3 rather than built from a new HUMAIN base. HUMAIN reports an 89.37% average across seven Arabic benchmarks, but the result comes from its own preview evaluation.
humain-m3 activates 23 billion parameters per token and was further trained on more than one trillion Arabic-native tokens.
HUMAIN’s reported 89.37% average exceeded MiniMax M3’s 80.34% reference and led five of seven tests.
Research-preview access includes thinking, streaming, tool use, computer use, and multimodal text, image, and video capabilities.
HUMAIN has put humain-m3, an Arabic-focused AI model built on the MiniMax-M3 lineage, into research preview through HUMAIN Node. The release gives developers a model further trained on more than one trillion Arabic-native tokens, but its eventual downloadable weights are still a future target rather than a present product.
Announced at LEAP in Riyadh, the model is a 428-billion-parameter mixture-of-experts system commissioned by HUMAIN and delivered by MiniMax. Mixture-of-experts models contain many parameter groups but select only part of the system for a given token; HUMAIN says humain-m3 activates 23 billion parameters per token. The company also says it further pre-trained the model on Arabic-native material after its MiniMax-M3 starting point.
The MiniMax foundation distinguishes the release from a wholly new base architecture developed by HUMAIN. HUMAIN calls the system a model built on the MiniMax-M3 lineage, while Crypto Briefing characterizes the underlying M3 weights as MiniMax open weights. Both descriptions point to a model adapted for Arabic from an existing MiniMax base.
HUMAIN reported an 89.37% average across seven equally weighted public Arabic benchmarks, ahead of the MiniMax M3 reference checkpoint at 80.34%. It also placed humain-m3 ahead of GPT-5.6 SOL at 87.30% and Opus 5 at 87.34%, and said the model led five of the seven tests. Those scores are HUMAIN’s own evaluation of a preview checkpoint, not a general verdict on all Arabic-language work.
The test set spans Arabic understanding, native and translated knowledge, academic exams, language proficiency, truthfulness, and retrieval-augmented generation. That breadth makes the single average useful, but prospective users will still need to judge the task most relevant to their deployment rather than treating the aggregate score as a substitute for application testing.
For now, the practical product is access through HUMAIN Node rather than a weight download. Developers can use a no-code playground or an OpenAI-compatible API. HUMAIN says its limited-preview tier applies a Saudi alignment guardrail, disables thinking and streaming, and adds latency; the research-preview tier offers the full checkpoint with thinking and streaming enabled and lower latency.
HUMAIN expects to release the weights under the MiniMax Community License after completing safety training and alignment, with a target of next month. It says the preview is meant to gather feedback on capability, safety, and alignment across Arabic dialects before general availability. The near-term test is whether HUMAIN meets that conditional target for a licensed weight release.
The model also joins HUMAIN’s stated portfolio alongside its ALLAM family of Arabic models. humain-m3 makes the company’s immediate proposition more concrete: an Arabic-adapted, large-scale model available for evaluation today, with openness and broader availability still dependent on work HUMAIN says is not yet complete.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.