Model details
ALLaM-2-7b is a bilingual Arabic-English language model built by the National Center for Artificial Intelligence at the Saudi Data and AI Authority, designed to advance Arabic language technology while preserving strong English ability. Its originators trained it from scratch in two stages, beginning with a large English-focused pass before continuing on a mixed Arabic and English corpus, a recipe aimed at transferring knowledge across languages without losing prior capability. The result is an instruction-tuned model positioned for conversational applications in both Arabic and English contexts.
On Groq, the model is delivered with a reported throughput of roughly 1,800 tokens per second and supports structured output through JSON Object Mode, making it a practical choice for developers building bilingual assistants or retrieval pipelines that need reliable, schema-aware responses. Its autoregressive design, paired with Groq's TruePoint numerics for efficient inference, gives it a useful balance of language coverage and responsiveness for chat-oriented workloads where Arabic fluency and English competence both matter.
A provider subscription or plan supersedes token-based pricing for this model.
Groq
As part of Saudi Arabia's drive towards AI leadership, SDAIA partners with Microsoft to bring ALLaM-2-7B to the Azure platform, enabling groundbreaking advancements in Arabic language technology and AI development.. ALLaM-2-7B: The Saudi AI Powerhouse Now Accessible via Microsoft's Azure. ALLAM, Arabic LLM, Saudi. AI &
Groq
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.
Groq
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.
