HUMAIN launches M3 open-weights Arabic model and a multi-dialect voice platform

HUMAIN used LEAP 2026 in Riyadh to unveil two new pieces of its Arabic-language AI stack: HUMAIN M3, an open-weights language model the company says was trained on more than one trillion Arabic tokens, and HUMAIN Voice, a conversational AI platform built to handle multiple spoken Arabic dialects rather than the standardised written form most language models default to. The dialect list matters more than it might look. HUMAIN Voice is designed to work across Saudi, Maghrebi, Egyptian and Levantine Arabic — varieties that differ from each other, and from Modern Standard Arabic, in vocabulary, pronunciation and grammar to a degree that a model trained mainly on formal written text tends to handle badly. A voice assistant that only understands the Arabic of news broadcasts is of limited use to someone speaking the dialect they use at home, and dialect coverage has been one of the harder unsolved problems in Arabic natural language processing. The open-weights framing on M3 is the other notable choice. Releasing a model's weights, rather than only offering it through an API, lets developers download, inspect, fine-tune and run it on their own infrastructure — a distinction that matters to government agencies and regulated industries wary of sending data to a hosted service outside their control. It is also a competitive statement: an open-weights Arabic model trained at the scale HUMAIN describes invites comparison with the smaller number of Arabic-focused open models built to date, most of which have trained on far less Arabic-specific text. Both launches sit inside a broader pattern for HUMAIN, which has spent LEAP week building out an Arabic AI portfolio on several fronts at once. Microsoft has agreed to bring HUMAIN's existing ALLaM models to Foundry and Microsoft 365 Copilot , giving Saudi-developed models a distribution channel through a global platform. HUMAIN has also been building the hardware to run models like M3 on, unveiling a Qualcomm-based AI PC under its own brand and lining up compute capacity through deals such as the AI Zone it is building with AWS . M3 and Voice are the model layer that portfolio has been assembled to run. What is not yet published is any independent benchmark for M3 against other Arabic or multilingual models, a release date or licensing terms for the open weights, or details of how HUMAIN Voice will be distributed — as a standalone product, an API, or a feature folded into HUMAIN's existing consumer devices. The one-trillion-token figure and the four-dialect list come from HUMAIN's own announcement rather than independent testing, and should be read as the company's claim about its training data and coverage rather than a verified capability. The more interesting test than the token count is whether M3 and Voice ship as usable products developers can build on, on a timeline HUMAIN is willing to commit to, given the run of hardware and infrastructure launches the company has made this week without matching detail on availability. TechScoop is reporting from LEAP 2026 in Riyadh.