HUMAIN's Arabic-first language model family
A primary-source deep dive into the four publicly documented ALLaM variants, the architecture, the training footprint, and every access channel we could verify.
ALLaM is the Arabic-first large language model family developed and commercialized by HUMAIN, Saudi Arabia's national AI champion, a Public Investment Fund company launched on 12 May 2025. The program is grounded in peer-reviewed research (the ICLR 2025 paper "ALLaM: Large Language Models for Arabic and English" by Bari et al.) and has shipped publicly documented variants on Hugging Face, IBM watsonx.ai, Microsoft Azure AI Foundry, and HUMAIN's own consumer chat product. The research and early variants originated inside the Saudi Data and Artificial Intelligence Authority (SDAIA) National Center for AI, with HUMAIN now serving as the commercial home for the family.
We track 4 ALLaM variants on LM Market Cap today. ALLaM is a purpose-built Arabic-first family: the program was designed around Arabic pretraining data, Arabic alignment, and Arabic evaluation from the start. Our composite leaderboard score is driven by English-language frontier benchmarks (MMLU, GPQA, HumanEval, SWE-bench), so it does not capture the dimension ALLaM was built to serve. The right way to read ALLaM is as a specialist model family anchoring an Arabic-language AI stack for Saudi Arabia and the broader Arabic-speaking world. Every claim in this report links back to a primary source so readers at HUMAIN can verify it directly.
What ALLaM Is
The peer-reviewed paper, the research origins, and the HUMAIN commercial handoff.
The ALLaM program is described in the paper "ALLaM: Large Language Models for Arabic and English" by M. Saiful Bari and the ALLaM team, first posted to arXiv on 22 July 2024 and accepted to ICLR 2025. The paper introduces four variants: a 7B model continued-pretrained from Llama 2, a 13B model, a 70B model, and a 7B model trained from scratch. The full paper is available at arxiv.org/abs/2407.15390.
HUMAIN is the AI company that now carries the ALLaM family forward. HUMAIN was announced on 12 May 2025 as a Public Investment Fund (PIF) company, chaired by HRH Crown Prince Mohammed bin Salman, with CEO Tareq Amin. It is positioned as Saudi Arabia's national champion for AI, data center, and foundation model work. ALLaM-branded model pages on Hugging Face are now mirrored under both the ALLaM-AI and humain-ai organizations.
The research origins of ALLaM sit inside the National Center for AI (NCAI) at the Saudi Data and Artificial Intelligence Authority (SDAIA), which published the ICLR paper and the early model cards. With HUMAIN now serving as the commercial home for the family, HUMAIN and NCAI together form the research-to-production pipeline behind ALLaM.
The Model Family
Four publicly documented variants across open weights, IBM watsonx, Azure AI Foundry, and HUMAIN Chat.
Four ALLaM variants have verifiable public footprints as of April 2026. The 70B model discussed in the paper has not been published with open weights and is not listed on a public commercial endpoint we could identify, so we do not track it here.
Hugging Face · Research license
The 7B Instruct preview is the most open, most studied variant. It is published on Hugging Face as ALLaM-AI/ALLaM-7B-Instruct-preview and mirrored byte-identically under humain-ai/ALLaM-7B-Instruct-preview. According to the Hugging Face model card, the publicly released preview is trained from scratch in two stages: 4T English tokens followed by 1.2T mixed Arabic/English tokens, then instruction tuned on curated Arabic and English data.
Managed API · Closed weights
The 13B Instruct tier is hosted as a foundation model on IBM watsonx.ai under the model id sdaia/allam-1-13b-instruct. It is listed in IBM's watsonx foundation model catalog at a published list price of roughly $1.80 per million tokens, applied to both input and output. IBM is a first-party commercial launch partner and the watsonx listing is the only public, metered API endpoint for any ALLaM variant we were able to verify.
Managed endpoint · Closed weights
ALLaM 2 is the second generation 7B Instruct model. It is listed in the Microsoft Azure AI Foundry model catalog as ALLaM-2-7b-instruct and can be deployed as a managed endpoint with pay-per-token billing. Azure pricing varies by region and SKU; we link to the Azure catalog rather than quoting a point-in-time number.
Consumer · Closed weights · No public API
The 34B variant is served exclusively through the HUMAIN Chat consumer product at chat.humain.ai. HUMAIN Chat launched on 25 August 2025 as the public-facing surface for the ALLaM family. Anyone wishing to try this variant today does so through the chat product.
Architecture
Every number here is pulled directly from the config.json on Hugging Face, not inferred.
The 7B Instruct preview ships with a config.json in its Hugging Face repository that spells out the architecture. These are the numbers that a HUMAIN engineer can confirm by opening the file directly.
Two architectural decisions are worth noting. First, ALLaM 7B runs with 32 full attention heads in a standard Llama-family configuration. Second, the vocabulary was expanded to 64,000 tokens (roughly 2x Llama 2's 32,000) to accommodate Arabic morphology and script. This is a deliberate design choice for an Arabic-first model: a larger Arabic-aware vocabulary gives the model more efficient tokenization on Arabic text, which improves both training efficiency and inference quality on Arabic inputs.
The 4,096 token context window is well sized for the Arabic instruction-following, dialogue, and question-answering workloads the 7B variant was targeted at. For longer-context Arabic workloads, the 13B tier on IBM watsonx and the 34B variant served through HUMAIN Chat offer additional headroom within the ALLaM family.
Training Data and Compute
The ICLR paper discloses the token budget and compute footprint in unusual detail.
The ALLaM paper describes the training footprint for the model family in unusual detail for a national-lab release. Section 3 of the paper gives the token budget and compute cost; the model card for the 7B Instruct preview published on Hugging Face states the release was trained from scratch in two stages.
For context, Meta has stated that Llama-2-70B took about 1.7M A100 GPU-hours, so the ALLaM program's cumulative ~5M A100-hour budget across four model sizes is a substantial but not exotic commitment. The paper explicitly positions this as a choice: the team was not trying to push frontier scale, they were trying to produce a high-quality Arabic-first model within a fixed, disclosed compute budget.
Benchmark Performance
Arabic-first evaluation, labeled as self-reported or independent. Read ALLaM on its home turf.
ALLaM is evaluated primarily on Arabic benchmarks. Our LM Market Cap composite is driven by English-language frontier evaluations (MMLU, GPQA, HumanEval, SWE-bench, Terminal-Bench, BrowseComp, etc.), so the composite score on the coding leaderboard does not reflect what ALLaM was built to do. The fair way to read ALLaM is on its home turf: Arabic language understanding, dialogue, and knowledge. This section shows that picture.
Self-reported · 7B Instruct preview
Independent evaluation signal
AraLingBench is the cleanest independent datapoint we found and places ALLaM 7B Instruct in the upper cluster of Arabic-capable open models on that specific benchmark. It aligns with the self-reported picture: Arabic linguistic competence and Arabic knowledge recall are clear strengths of the family.
Because ALLaM is an Arabic-first specialist, it is not the target of the major independent English-language LLM leaderboards (Chatbot Arena, LiveBench, SEAL, MMLU-Pro, HLE), and we did not find it listed in their top tiers as of April 2026. We also checked the public Arabic leaderboards Scale SEAL Arabic, OALL (Open Arabic LLM Leaderboard), and AraGen. ALLaM is not currently listed in the public top of these specific leaderboards either; this is a gap in independent ranking data rather than a statement about model quality, and it reflects that ALLaM's primary eval harness is the one published with the ICLR 2025 paper.
Where ALLaM Is Available
Four publicly documented access channels across open weights, two clouds, and HUMAIN's own chat product.
Four access channels are publicly documented. They serve very different audiences.
Hugging Face
Open weightsALLaM-AI/ALLaM-7B-Instruct-preview on Hugging Face. Open weights under a research license. The right channel for academics, Arabic NLP researchers, and anyone who wants to inspect or fine-tune the model.
IBM watsonx.ai
Managed APIsdaia/allam-1-13b-instruct as a managed foundation model. Roughly $1.80 per million tokens on both input and output. Good fit if you are already on IBM watsonx and need an Arabic-tuned model.
Microsoft Azure AI Foundry
Managed endpointALLaM-2-7b-instruct as a managed endpoint in the Azure AI Foundry catalog. Pricing varies by region. Good fit if you are already on Azure and need an Arabic-tuned option in the same catalog as GPT-5, Llama, and Mistral.
HUMAIN Chat
Consumerchat.humain.ai/en is the only way to interact with ALLaM 34B. Free, web-based, public-facing. Not a developer API.
The Saudi National AI Program
ALLaM is the flagship language model of a national program that has been building in public since 2019.
ALLaM is best read as the flagship language model of a national AI program that Saudi Arabia has been building in public since 2019. The timeline, the institutions, and the numbers are all on the record, and they explain how a national-scale Arabic model ended up with a peer-reviewed paper, four public variants, and commercial endpoints on two major clouds within a few years.
HUMAIN is the company that now carries the ALLaM family forward commercially. The research lineage runs through the National Center for AI (NCAI), the NSDAI strategy document ties the program to Vision 2030, and the published Vision 2030 targets are the ones the broader national AI program is measured against.
HUMAIN and the Saudi AI Ecosystem
HUMAIN's publicly announced infrastructure and model partnerships, sourced from primary press releases.
HUMAIN is the commercial layer and the engine that gives the ALLaM family its path to scale. As of late 2025 and into 2026, the company has been assembling the largest publicly-announced set of AI infrastructure partnerships of any company in the region. The deals are public; the scale is unusual. We list only the partnerships that we could verify from a primary HUMAIN or partner press release.
Announced 18,000 Blackwell Ultra GB300 GPUs for HUMAIN, positioned as the initial tranche of a multi-year commitment that both companies have publicly discussed scaling into the hundreds of thousands of GPUs.
AWS announced a $5B+ commitment to build a dedicated AI Zone in Saudi Arabia in partnership with HUMAIN, covering compute, training services, and support for Saudi AI startups.
Google Cloud and PIF announced a joint $10B commitment to build an AI hub in Saudi Arabia. HUMAIN is the PIF-side counterparty for AI workloads.
Groq committed to day-zero availability for OpenAI's gpt-oss open-weights release on HUMAIN's Groq-powered inference stack, giving HUMAIN a high-speed inference option independent of its own training runs.
Qualcomm announced a partnership with HUMAIN to deploy AI200 and AI250 chips across 200 MW of data center capacity in Saudi Arabia, giving HUMAIN a second-source inference path alongside NVIDIA.
Adobe and HUMAIN announced a partnership around Creative Cloud and Firefly deployments into the Saudi market, with Arabic-language localization as a focus area.
xAI and HUMAIN announced an agreement to make Grok available through HUMAIN's distribution channels, positioning HUMAIN as a multi-model distributor inside the kingdom.
Aramco announced a minority stake in HUMAIN in October 2025, extending ownership beyond PIF alone and tying HUMAIN into the Aramco enterprise customer base.
The strategic picture is that ALLaM and the HUMAIN infrastructure program are being built in lockstep. Saudi Arabia is assembling frontier-scale AI infrastructure in parallel with its own Arabic-first model program, and the two reinforce each other: ALLaM anchors the Arabic language capability, and the HUMAIN data center and GPU footprint gives that capability a path to scale. The 2025-2026 releases are the opening chapter of a multi-year program, and the 18,000 GB300 tranche, the AWS AI Zone, and the Google Cloud / PIF hub make the longer trajectory clear.
Transparency Notes
Open items we chose to leave out of the main body until we can anchor them to a primary source.
We tried to source every claim above from a primary document. The items below are things we chose not to include in the main body because we could not anchor them to a single stable public reference. They are listed here as open items to update, not as criticism. If HUMAIN publishes new primary sources, we will fold them into the report.
How to Use ALLaM Today
A practical evaluation and deployment path for Arabic-first teams, as of April 2026.
If you are building an Arabic-first product and want to evaluate ALLaM, this is the practical path as of April 2026:
ALLaM on LM Market Cap
Composite scores for cross-category comparison; model-specific pages for Arabic benchmark detail.
We track all four publicly-documented ALLaM variants on LM Market Cap. Our composite score is driven by English-language frontier benchmarks, so it measures a different axis than the Arabic-first evaluation ALLaM was designed for. Use the composite for cross-category comparison on context window, pricing, and capability flags, and use the model-specific pages below for Arabic benchmark detail, access channels, and primary-source links.
Primary Sources
Every claim in this report links to one of the primary documents below.