0.9B, 3.7B, 7B, and 32B dense; 36B total / 4B active (MoE + MoVA); 375B total / 23B active (MoE)
512K tokens (128K for the 0.9B model)
Apache License 2.0
Six-model open-weight family focused on Arabic-native reasoning, instruction following, and agentic work while retaining English and general reasoning capabilities from the underlying K2 Horizon models. Jais 3 covers Modern Standard Arabic, Emirati and Egyptian dialects, Arabizi, Arabic poetry and cultural knowledge, and is trained to plan multi-step tasks, call tools and MCP servers, search the web, run terminal commands, and edit Arabic documents, presentations, and spreadsheets. The family comprises four dense checkpoints at 0.9B, 3.7B, 7B, and 32B; a 36B-A4B sparse model combining Mixture-of-Experts layers with Mixture-of-Values Attention; and a 375B-A23B flagship MoE. Every model except the 0.9B supports a 512K-token context window; the 0.9B supports 128K. IFM released final weights, intermediate checkpoints from each training stage, configs, recipes, logs, sample training data, audited Arabic benchmarks, and evaluation code. Training includes Arabic continued pretraining, supervised agent data, Final Token Preference Optimization for reducing repetitive reasoning loops, and verifiable-reward reinforcement learning across instruction following, Arabic grammatical analysis, mathematics, diacritization, and code. IFM reports AA-LCR scores of 65 for Jais 3 7B versus 49 for Gemma 4 12B, and 75 for Jais 3 32B versus 66 for Gemma 4 31B. Optional rank-128 Uno adapters are available for the four dense models, converting them to discrete-diffusion generation with speculative sampling; the Uno release reports an average 2.15x H100 speedup while targeting the original model distribution. BF16, FP8, GGUF, and MLX variants are available for supported sizes.