ALLaM-2-7b

ALLaM-2-7B-Instruct is a 7-billion-parameter autoregressive transformer trained from scratch by the Saudi Data and AI Authority (SDAIA), featuring a 4,096-token context window and optimized for low-resource deployment. The model is part of the ALLaM family's second generation, which moved away from continued pretraining of Llama 2 weights in favor of a fully native architecture. Its release reflects Saudi Arabia's designation of 2026 as the "Year of AI" and its broader push to anchor an Arabic-first foundation model stack. The model name, meaning "worldly-wise," signals its mandate to serve as the definitive intelligence layer for the Arabic-speaking world. ALLaM-2-7B-Instruct was open-sourced on Hugging Face under an Apache 2.0 license, transitioning ALLaM from a closed-API national project into a globally accessible open-weight ecosystem. This open-weight milestone anchors a larger expansion alongside a 1.8-trillion-parameter enterprise model backed by HUMAIN. The release positions ALLaM to compete with open-source leaders like Llama 3 and proprietary systems like GPT-4o in Arabic-language tasks. SDAIA's tiered ecosystem approach combines this accessible 7B variant with a much larger enterprise-tier model for sovereign AI deployment.

ALLaM-2-7B is a bilingual Arabic-English language model developed by the National Center for Artificial Intelligence (NCAI) at the Saudi Data and AI Authority (SDAIA) and hosted on Groq, which lists it under the allam-2-7b endpoint with a 4,096-token context window and up to 4,096 max output tokens. The model is an autoregressive transformer with 7 billion parameters, trained from scratch using a two-step recipe of 4T English tokens followed by 1.2T mixed Arabic/English tokens, designed to advance Arabic Language Technology. Groq reports benchmark results including MMLU English (0-shot) at 63.65%, Arabic MMLU at 69.15%, ETEC Arabic at 67.0%, IEN-MCQ at 90.8%, MT-bench Arabic Average 6.6/10, and MT-bench English Average 7.14/10. The endpoint runs at approximately 1,800 tokens per second using Groq's TruePoint Numerics quantization, supports a JSON Object Mode capability, and is governed by Humain's Terms of Service rather than Groq's standard policy.