MBZUAI Releases Jais 3: Six Open Arabic Models from 0.9B to 375B with Full Training Recipe
lmoroney · x · 2026-10-19
MBZUAI's Institute of Foundation Models, with Inception and Cerebras, released Jais 3 — six open Arabic-centric models ranging from 0.9B to 375B parameters.
The training recipe is instructive for anyone building models in other languages: each model starts from the matching K2 Horizon checkpoint and goes through continued pre-training, a cool-down phase, supervised fine-tuning, a step targeting repetition loops, and reinforcement learning.
Unusually open release: weights, checkpoints from every training stage, configs, training logs, and eval code are all public under Apache 2.0. The author suggests picking an intermediate checkpoint and running your own evals against the final model to see what each stage contributed.
Related event: MBZUAI Open-Sources Jais 3 Arabic Model Family Up to 375B Parameters(2 posts)→
More from Models
- MIST Framework Separates Execution from Memory to Diagnose Video World Model Failures — hkuhk · 2026-10-19
- Reddit test finds Nemotron 3.5 Super VL best at obscure knowledge, beating GPT-OSS and Qwen — Kahvana · 2026-10-19
- Thought experiment: if billion-dollar companies stopped open-weight releases for a year — ForsookComparison · 2026-10-19
- Opus 5.5 hailed as a milestone: developers say they've stopped chasing new models — dreamwieber · 2026-10-19
- Six frontier AIs design egg-drop capsules; Claude Opus 5.5 survives 20 feet — 141_1337 · 2026-10-19
- Google AI Pro users from Pixel/Jio/student offers locked out of Gemini 4 Argon in Antigravity — MicahBerkley · 2026-10-19