Jais 3 375B-A23B

Jais 3 375B-A23B is the flagship of the first model family on this site built for Arabic rather than scoring on it. The Institute of Foundation Models, part of Mohamed bin Zayed University of Artificial Intelligence, published it with Inception42 on October 15, 2026 as the largest of six open models spanning 0.9B to 375B, alongside an Arabic speech recognition model called Jais-ASR. It is a mixture-of-experts with 375 billion total parameters and 23 billion active per token, around 758GB in BF16, which puts it on one 8xH200 node, with a 512,000 token context window and Apache 2.0 weights in the IFM collection on Hugging Face. There is no API price here and that is not an omission: IFM publishes weights and a self-hosting path, not a hosted endpoint, so the only cost is the hardware you run it on. The claims are worth reading for what they do and do not say. IFM reports an Arabic average of 84.0 against 82.0 for Qwen3.5-397B and says it also clears Claude Sonnet 5, while GPT-5.6 Sol and Gemini 3.8 Flash stay ahead; on its reported agentic aggregate it comes second among large open models behind Qwen3.8-Flash-Next, and on an Arabic factuality aggregate built from SimpleQA and HalluScore it scores 51.0. It trails Gemini and GPT-5.6 on AA-LCR, an Arabic long-context reasoning benchmark. Every figure is IFM-reported, which matters less than usual because the group also published intermediate checkpoints, training recipes, logs, and a cleaned Arabic benchmark suite, so the runs are reproducible by anyone with the hardware. The smaller siblings are the practical story for most teams: the 7B model reports 79.4 on Arabic against 66.7 for the much larger Jais 2 70B, and the 0.9B model fits in roughly 2.2GB in BF16 for edge use with a 128K window. Infrastructure came from Inception and Cerebras Systems.