MBZUAI's Jais 2 brings 70B Arabic-centric LLM with 2000 tokens/sec
Largest open Arabic-centric LLM at 70B beats benchmarks with efficient training
MBZUAI, Cerebras, and Inception have unveiled Jais 2, a family of Arabic-centric large language models designed to push the boundaries of Arabic-language AI. The family includes a 70B-parameter model, which the team says is the largest open Arabic-centric LLM trained from scratch, alongside a compact 8B variant that remains competitive among open models. A custom Arabic-centric vocabulary, combined with an optimized architecture and training recipe, delivers highly compute-efficient training. As a result, Jais 2 uses a substantially smaller token budget than comparable models while achieving strong Arabic performance and competitive English results.
Jais 2 leads among evaluated open models on the OALL2 and AraGen benchmarks, and performs strongly on culturally grounded Arabic tasks like poetry, religion, cuisine, and dream interpretation, as well as general tasks such as translation and summarization. The models are released on HuggingFace under a commercially permissive license. The 70B version is also available as a chat app on Web, iOS, and Android, running on Cerebras hardware to deliver up to 2,000 tokens per second for high-throughput serving. By combining scale, linguistic diversity, cultural fidelity, openness, and speed, Jais 2 provides an open-weight foundation for advancing Arabic-centric NLP research and development.
- Jais 2 includes the largest open Arabic-centric LLM trained from scratch at 70B parameters, plus an 8B variant
- Delivers up to 2,000 tokens per second on Cerebras hardware, enabling high-throughput Arabic chat
- Leads open models on OALL2 and AraGen benchmarks, with strong scores on poetry, religion, and cuisine tasks
Why It Matters
Jais 2 gives researchers and enterprises an open, fast, commercially usable foundation for Arabic AI applications.