Jais 3 ships six open Arabic models with a 512K-token context
IFM and Inception42 release an open Arabic-native model family from 900 million to 375 billion parameters, plus a dedicated speech model, with weights and checkpoints on Hugging Face.
The Institute of Foundation Models (IFM) and Inception42 released Jais 3 on 16 October: six open Arabic-language models ranging from 900 million to 375 billion parameters, plus Jais Speech, a dedicated Arabic speech recognition model. All are built on K2 Horizon, IFM’s open foundation model family from September, adapted to a new language for the first time.
Every Jais 3 model shares a 512,000-token context window, which IFM says is the longest of any Arabic-native model. At the 7-billion-parameter size, IFM’s own Arabic evaluation score rose from 56.4 for the base K2 model to 79.6 for Jais 3. Jais Speech transcribes Modern Standard Arabic and regional dialects across 19 Arab countries, handles code-switching into English or French, and adds the vowel marks most transcription tools drop. Weights, benchmark results and five training checkpoints per model are on Hugging Face, with deployment support for vLLM and SGLang.
Why it matters: open-weight releases overwhelmingly target English first. A genuinely open Arabic-native stack, speech model included, with checkpoints and benchmarks published alongside it, gives anyone building Arabic-language products a self-hostable alternative to commercial APIs rather than a translation layer bolted onto an English model.
The caveat: the benchmark numbers are IFM’s own. Independent evaluation of the performance claims hasn’t happened yet.