IFM Releases Jais 3 Arabic Model Family with Staged Weights and Training Recipes

The Institute of Foundation Models (IFM), part of MBZUAI, released Jais 3 on October 15, extending the goals of Arabic models to document editing, search, and multi-step tool use. The family is based on K2 Horizon and includes four dense models and two sparse models, 36B-A4B and 375B-A23B. The models and code are released under the Apache 2.0 license. Official announcement

The language adaptation process is a notable technical aspect. Starting from post-trained base models, the team conducted continued pretraining, cooldown training, supervised fine-tuning, preference optimization, and reinforcement learning, in that order. Continued pretraining mixes Arabic data with data from the original model to build capability in the new language while retaining existing knowledge. Agent capabilities come mainly from tool-use trajectories in supervised fine-tuning; the final reinforcement learning stage did not use agent tasks. This allows researchers to examine separately where language adaptation and tool-use capabilities come from. Training process

For deployment, all models except the smallest are listed as supporting a 512K context; the smallest supports 128K. The 7B model card provides launch instructions for vLLM and SGLang. Both inference and tool calling use the k2_horizon parser. Tool arguments can use JSON, XML, or typed XML formats. Engineers need to configure the chat template and parser together to connect model outputs to a production tool-execution loop. 7B model card

Reproducibility is a significant part of this release. The model card lists weight branches for continued pretraining, cooldown, supervised fine-tuning, and preference optimization, and links to training records. The official collection also provides quantized, GGUF, and MLX versions for Apple silicon. However, the 7B model card still marks the technical report and code repository as “coming soon,” so the availability of the complete research materials should be verified item by item. Release assets, model collection

The evaluation methodology warrants attention: the official team cleaned some Arabic test questions and switched to scoring generated answers, so the results cannot be directly compared with scores on older leaderboards. The model card also notes that the reported results use high reasoning effort. For Chinese technical teams, these staged assets may serve as a reference for language-adaptation research in other languages. Chinese-language capability, long-context memory requirements, and real-world agent latency still need to be tested independently. Evaluation notes, usage limitations