A sovereign model, openly licensed
On February 11, 2026, at the opening of the inaugural Ai Everything MEA summit in Cairo — held under the patronage of President Abdel Fattah el-Sisi — Egypt's Ministry of Communications and Information Technology (MCIT), working with the IT Industry Development Agency (ITIDA), unveiled Karnak, described as the country's first nationally developed large language model (ITIDA press release). Officials called it "a local intelligence foundation on which startups, enterprises and public institutions can build local AI solutions" — infrastructure, not a product.
The notable choice is how it was released. Karnak-40B, built by depth-extending Alibaba's open-weight Qwen3-30B, is published on Hugging Face under the Apache 2.0 license — a permissive open-source license that allows anyone, anywhere, to download the weights, run them locally, fine-tune them, and redeploy commercially with no royalty and no government approval (Karnak-40B-v1.0 model card). That is a meaningfully different posture than most "sovereign AI" projects, which tend to stay locked inside state clouds. Egypt's own strategy already frames applications built on Karnak — SIA, an AI tutor for high-school Arabic and history; a legal-navigation assistant; diagnostic tools for diabetic retinopathy and breast cancer — as proof points for a downloadable public model, not a walled-off one (MCIT, Artificial Intelligence).
The regulatory track running in parallel
The open release did not happen in a governance vacuum. Egypt's National Council for Artificial Intelligence (NCAI), established in 2019 and chaired by the ICT minister, has spent the past three years building out a binding framework: the 2023 Egyptian Charter for Responsible AI, the Egyptian Center for Responsible AI (ECRAI) stood up at the end of 2025, and a second-edition National AI Strategy (2025–2030) organized around six pillars — governance, technology, data, infrastructure, ecosystem, and talent (MCIT). The strategy sets concrete targets: AI and ICT lifting the sector's GDP contribution to 7.7% by 2030 and supporting more than 250 new AI startups (OECD.AI policy dashboard).
A draft AI law — still moving through stakeholder consultation as of mid-2026 — is expected to convert much of the Charter's voluntary language on transparency, fairness, and risk management into enforceable obligations, likely tiered by risk level, with ECRAI empowered to audit and certify compliance. Egyptian officials have leaned into the framing that AI is now "a strategic enabler influencing national security, economic development and state capabilities" rather than a narrow technology question (Middle East Observer).
The case for caution, stated fairly
That framing has real merit, and it shouldn't be waved away. Karnak isn't a toy — SIA is already reaching teenagers studying for exams, the legal assistant is meant to help small businesses navigate regulation they may not otherwise understand, and the healthcare models touch diagnostic decisions. A government that ships tools like these into classrooms, courts, and clinics, then declines to define who is liable when a diagnostic model errs or a legal assistant misstates the law, is ducking a real responsibility. Risk-tiered obligations, audit rights for a technical body like ECRAI, and clear routes to redress for affected citizens are not bureaucratic overreach in that context — they are what distinguishes a serious governance framework from a press release. Regulators worldwide, from the EU's AI Act to emerging frameworks across the Gulf, have converged on some version of this instinct for good reason.
Why the open license is the part worth protecting
Where the balance tips is in how those obligations get attached. An Apache-licensed base model is, by design, something anyone can pick up and modify without asking Egypt's permission — that is precisely why Karnak can plausibly seed the 250-plus startups the government says it wants. If the coming AI law imposes compliance costs — mandatory audits, certification fees, disclosure regimes calibrated for large deployers — at the level of anyone who fine-tunes an open model, rather than at the level of whoever actually deploys a high-stakes application (a diagnostic tool, a legal assistant), it will quietly punish the small Egyptian developers the strategy is meant to cultivate while leaving well-resourced incumbents unaffected. That outcome would be a regulatory own-goal against a strategy still in its first year.
The fix is not to relax scrutiny of high-risk deployments — it's to keep the regulatory hook on deployment context rather than on model provenance. A student chatbot built on fine-tuned Karnak and a hospital diagnostic tool built on the same weights are not the same risk, and treating an open, freely-modifiable foundation model as inherently regulated — rather than the specific uses built on top of it — would undercut the one part of Egypt's AI push that other governments in the region have not matched. Egypt built something genuinely open. The forthcoming law's job is to regulate what people do with it, not whether they may have it at all.
Officials describe Karnak as "a local intelligence foundation on which startups, enterprises and public institutions can build local AI solutions" — a description that only holds if the license underneath it stays open once the law arrives.