Dicta

A nonprofit, not a company: Prof. Moshe Koppel and the mission since 2015

Dicta is not a commercial company but a registered Israeli nonprofit, set up in 2015 on the initiative of Prof. Moshe Koppel. He is a professor emeritus of computer science at Bar-Ilan University, holds a doctorate in mathematics, and is a Talmud scholar in his own right. Koppel defined the organization's goal as removing the drudgery from the study of classical and modern Hebrew texts, so that researchers can focus on the more essential questions. In an interview with The Jerusalem Post, Koppel said: "Everything we do is 100% free." All of its work is funded by donations alone, with no business model, product sales or commercial fundraising of any kind. Dicta is based in Jerusalem and, according to published figures, employs about 49 people alongside seven members of the association.

Nakdan: solving the problem of missing vowel marks in Hebrew

Dicta's oldest and most central tool is Nakdan, a system for automatically adding vowel marks to Hebrew text. It was developed by Dr. Avi Shmidman, Shaltiel Shmidman, Prof. Moshe Koppel and Prof. Yoav Goldberg. The problem it solves is unique to Hebrew: because Hebrew text is usually written without vowel marks, the exact same string of letters can be read in several different ways depending on the context of the sentence. Nakdan does not rely only on a fixed word list but on neural networks that learn to understand the context and choose the correct vowels from it. This technology was built over five years of development and is the basis for several of Dicta's more advanced tools, including vowel marking for biblical, Talmudic and rabbinic texts.

From the vocalized Talmud to the digital library

Using its vowel-marking technology, Dicta produced a complete vocalized version of the Babylonian Talmud. This is a huge project, since marking thousands of pages of complex Talmudic text by hand had until then been a task that required decades of work by human experts. Beyond that, Dicta built a digital library that, according to published figures, includes about a thousand translated and digitally processed sacred texts. They are offered to the public under an open license (CC-BY-SA 4.0) that allows reuse. The library also includes advanced search engines for biblical, Talmudic and rabbinic texts, with context-based search and optical character recognition (OCR) for scanned manuscripts. These tools turn material that used to be accessible only to experts into an open resource for the general public.

Dicta-LM: an open-source Hebrew language model, with MAFAT and Nvidia

In 2024 Dicta launched Dicta-LM 2.0, a large language model based on the Mistral-7B architecture, with a vocabulary of tokens (a tokenizer) expanded especially for Hebrew to process the language more efficiently. The model was released as open source under the Apache 2.0 license, in collaboration with MAFAT (the Israeli Defense Ministry's Directorate of Defense Research and Development), the Israel Association for Language Technologies, and researchers from Intel. Later the organization also developed a new generation, Dicta-LM 3.0, a more advanced Hebrew reasoning model, this time in collaboration with NVIDIA. This is part of a wider trend in the industry to develop what is called "sovereign AI" for languages that do not get enough support from the largest global models. The ways such models are tested, including the open leaderboard for Hebrew language models, the HeQ dataset and blind human judging, are explained in the Hebrew entry on Evaluating Hebrew Language Models.

Rav Dicta: AI that answers questions of Jewish law

Beyond its text analysis tools, in March 2025 Dicta released a beta version of "Rav Dicta" (Rabbi Dicta), an AI model that answers questions of Jewish law (halakha) in the style of responsa, that is, in the question-and-answer format known from halakhic literature, based on the classical rabbinic literature the system was trained on. This tool is different in nature from everything else Dicta develops. Vowel marking, search and manuscript recognition are text-processing tasks with a correct answer you can check against. A halakhic answer, by contrast, is a ruling. Here the system moves from helping a person read an existing text to producing text itself, in a field whose authority traditionally rests on an identified human decisor, someone you can ask a follow-up question, so that you know exactly whose authority you are relying on. Dicta released the system as a beta version, that is, as a declared experimental stage and not a finished product.