The world's frontier Arabic model — released by HUMAIN, available now in limited preview on HUMAIN Node.
The previewed HUMAIN M3 leads five of seven Arabic benchmarks against the strongest frontier models available today.
The suite spans what production Arabic actually demands: core understanding, native and translated knowledge, academic examinations, language proficiency, truthfulness, and retrieval-augmented generation.
HUMAIN M3 is also natively multimodal — it understands images and video alongside text.
Against the reference model it builds on, HUMAIN's Arabic post-training adds nine points on average — the difference between speaking Arabic and being built for it.
Bold marks the best score per benchmark. Scores are HUMAIN's evaluation of the previewed humain-m3 checkpoint; seven public Arabic benchmarks, equally weighted.
| HUMAIN M3 | GPT-5.6 SOL | Opus 5 | M3 reference | |
|---|---|---|---|---|
| Core Arabic understandingAlGhafa | 86.45% | 81.54% | 83.06% | 75.31% |
| Broad knowledge · native ArabicArabicMMLU | 90.70% | 88.82% | 88.77% | 81.80% |
| Academic examinationsArabic EXAMS | 67.67% | 66.40% | 64.80% | 61.27% |
| Arabic language proficiencyMadinahQA | 95.44% | 94.48% | 94.78% | 87.10% |
| Truthfulness & trustAraTrust | 97.53% | 93.42% | 91.38% | 90.60% |
| Retrieval-augmented generationALRAGE | 94.63% | 94.91% | 94.85% | 79.32% |
| Broad knowledge · translatedTranslated MMLU | 93.20% | 91.48% | 93.78% | 87.00% |
| AverageSeven benchmarks, equal weight | 89.37% | 87.30% | 87.34% | 80.34% |
HUMAIN M3 builds on a natively multimodal mixture-of-experts engineered for agents, with HUMAIN's Arabic post-training on top.
Frontier tool use and computer use, built for long‑horizon agent workflows in Arabic and English.
Three thinking modes — always‑on, adaptive, and off — so every request spends exactly the reasoning it needs.
Text, image, and video trained jointly from the first step, with long‑video understanding and native screen operation.
Use the previewed model in the playground, or call humain-m3 through an OpenAI-compatible API.
Chat with HUMAIN M3 — no code, no setup.
Playground • HUMAIN M3اشرح لي الحوسبة الكمّية في ثلاث جمل.
تعتمد الحوسبة الكمّية على الكيوبتات القادرة على تمثيل عدّة حالات في آنٍ واحد، ما يتيح معالجة متوازية هائلة…
Chat in Arabic and English
No code or setup required
Test Saudi dialects and alignmentIntegrate HUMAIN M3 into your apps or your coding CLI.
quickstart.py — humain-m3OpenAI-compatible SDKs
Connect your agentic coding CLIBuild it into your own appsThree steps take you from a new account to the full frontier model.
Sign up on HUMAIN Node. Your account is the front door to the platform — and to the 100+ models as access widens.
Once approved, you can use HUMAIN M3 with the Saudi alignment guardrail — made for everyday use and testing.
The guardrail adds a little latency, with thinking and streaming off.
Once approved, you can use the full checkpoint to push the model to its limits — and your feedback shapes the next checkpoints.
HUMAIN M3 is the first model open to preview users on HUMAIN Node. 100+ more — frontier, proprietary, and open source, from the world's leading providers — are already on the platform, with broader access opening soon.
One API, one key, one billEvery model that lands on Node works through the same endpoint, with usage and cost in one view.Budgets, access, and limitsSet spending limits by team and project, and control who can use which models.Global, in-Kingdom, or sovereign hostingChoose where each workload runs — without changing your code.The preview period lets us evaluate capabilities, safety, and alignment across Arabic dialects with early users before general availability.
HUMAIN M3
GPT-5.6 SOL
Opus 5
M3 reference
