QeRN Weekly — week 2026-W38
Five stories this week, and the thread running through them is the difference between a rule and a ruling. A lab publishes not just its model weights but the data and the method, so anyone can check the claim. Twenty-five education ministers sign a statement and — more usefully — publish the procurement briefs underneath it. A development bank prices its second sukuk of the year at a spread nobody finds remarkable. A university lab opens speech and handwriting models for twenty-two languages. And a fiqh council answers the AI question with six enumerated conditions rather than a verdict. In every case the work is in the specification, not the announcement.
This week on QeRN
Abu Dhabi's IFM ships six models with the training data attached
The Institute of Foundation Models, launched by MBZUAI in 2025, released K2 Horizon on 3 September: six models from 0.9B to 375B parameters under Apache 2.0. What ships alongside the weights is the point — training data, data recipes, training code, configurations, intermediate checkpoints, training logs and evaluations. Founder Eric Xing put the argument directly: open source is much more than open weights, and science works when others can see the data, follow the method, reproduce the result and improve on it. The fleet runs from a model small enough for a watch to a flagship built for enterprise reasoning, sharing one architecture and one deployment path across all six. Why it matters: reproducibility is what lets a university in Lahore or Kano audit a model rather than merely license one. A lab funded by a Muslim-majority state has made verifiability, not secrecy, its competitive position.
Twenty-five education ministers sign a statement on who governs AI in schools
UNESCO's Digital Learning Week ran in Paris from 8 to 11 September and closed with a joint ministerial statement on sustaining education as a common good in the age of AI. More than twenty-five ministers signed, with the text pressing on public accountability, the rights of learners and teachers, and stronger governance over how AI enters a school system. The durable output is the material published with it: policy briefs on AI safety governance, digital sovereignty, procurement, cost models and age-appropriateness. Procurement is the least glamorous item and the one that will decide most outcomes — a ministry that has not settled contract terms, data rights and exit costs before a vendor arrives will take whatever is offered. Why it matters: these are short documents written for officials who have to sign something next quarter, which makes them usable rather than aspirational.
The Islamic Development Bank prices a second billion-and-a-half sukuk on ordinary terms
On 2 September the IsDB priced US$1.5 billion of five-year trust certificates, its second benchmark sukuk of 2026, under a US$25 billion programme. The deal closed at 48 basis points over SOFR mid-swaps, two inside guidance, at par, with a 4.781 per cent profit rate paid semi-annually. The order book passed US$2.75 billion by midday London time, ahead of the bank's own previous high set in May. Proceeds fund projects across member countries. Why it matters: the spread is the story, not the size. Forty-eight over swaps is what a conventional triple-A multilateral would pay, which means Shariah-compliant paper is being bought on ordinary credit terms by ordinary institutional buyers. A second issuance matters more than a first, because repetition is what turns an instrument into infrastructure.
Open models for Indian languages arrive, and Urdu handwriting is in scope
Bodhan AI, incubated at IIT Madras, has released open foundational models for Indian languages built with NVIDIA and AI4Bharat. Four capabilities ship together: speech recognition across 26 Indian languages plus English, text-to-speech across 23, translation across 22, and OCR for printed and handwritten text. Urdu is covered, handwriting included — which should interest anyone who has tried to get a machine to read a nastaliq manuscript, a madrasa register or a family archive. The weights are open, with hosted APIs on sovereign infrastructure, and the education applications built on them are to remain free for learners, teachers and participating state governments. Why it matters: a language whose recognition models are open can be taught, archived and searched by whoever cares to. A language dependent on a vendor's roadmap waits its turn.
The Fiqh Academy's standing ruling on artificial intelligence, and its six conditions
The International Islamic Fiqh Academy of the OIC issued Resolution 258 (3/26) on AI at its twenty-sixth session in Doha, 4-8 May 2025. The default ruling on developing and using AI is permissibility, ibaahah, subject to six conditions: lawful purpose in creation, use, funding and outcomes; net benefit over harm; no insult to beliefs, religions or religious symbols; protection of information and of public and private rights; no threat to individual, societal or national security; and honesty, documentation and transparency in use. The resolution builds on the Academy's earlier rulings on moral rights and smart contracts, and on ICESCO's 2025 Riyadh Charter on AI for the Islamic World. It leaves legal personality for AI explicitly open, recommending study rather than declaring an answer. Why it matters: a default of permissibility with enumerated conditions gives a developer, a procurement officer or a school board something to check against. A blanket approval or a blanket warning gives them nothing.
If one of these is worth an hour of someone's week — a procurement officer, a registrar, a treasurer, a teacher — forward it to them. They can subscribe at https://qern.org/.