By HUMAIN, Saudi Arabia’s Public Investment Fund AI company, launched humain-m3 at LEAP 2026 in Riyadh on September 3. It is a 428-billion-parameter mixture-of-experts model built on China’s MiniMax-M3 base and post-trained on over a trillion tokens of Arabic content. The headline claim: an 89.37% average across seven public Arabic benchmarks, topping GPT-5.6 and Claude Opus 5. API access is open for requests today, and weights are targeted for October. If you build products for Arabic-speaking markets, this is the model to watch — with a few caveats worth knowing upfront.
Under the Hood
humain-m3 inherits MiniMax-M3’s architecture wholesale. That means 428 billion total parameters with 23 billion active per token — a mixture-of-experts design with 128 experts and 4 activated per token. The efficiency story is the MiniMax Sparse Attention (MSA) mechanism, which cuts per-token compute to roughly 1/20th of the previous generation at 1-million-token context. In practical terms: long documents, long conversations, and long-form Arabic content do not create the latency cliff you see with dense transformers at scale.
The model is natively multimodal — text, images, and video were trained together from step zero, not bolted on later. Three thinking modes are available (always-on, adaptive, and off), which matters for Arabic tasks that range from quick lookups to multi-step legal document analysis.
The Arabic-specific addition is the post-training corpus: over one trillion tokens of Arabic-native content, resulting in approximately 9 to 11 percentage point gains over the base MiniMax-M3 model on core Arabic benchmarks.
The Benchmark Numbers
HUMAIN reports humain-m3 leads on five of seven public Arabic benchmarks, with an 89.37% average against GPT-5.6’s 87.30% and Claude Opus 5’s 87.34%:
| Benchmark | humain-m3 | GPT-5.6 | Opus 5 |
|---|---|---|---|
| AlGhafa | 86.45% | — | — |
| ArabicMMLU | 90.70% | — | — |
| MadinahQA | 95.44% | — | — |
| AraTrust | 97.53% | — | — |
| ALRAGE | 94.63% | 94.91% | — |
| Translated MMLU | 93.20% | — | 93.78% |
| Arabic EXAMS | 67.67% | — | — |
| Average | 89.37% | 87.30% | 87.34% |
Two weak spots are worth flagging. The ALRAGE retrieval-augmented generation benchmark puts GPT-5.6 marginally ahead (94.91% vs 94.63%), which matters if you are building Arabic RAG pipelines. Arabic EXAMS at 67.67% is the most notable outlier — exam-style, academic Arabic is clearly not humain-m3’s strongest mode. Plan accordingly if your use case leans toward standardized test formats or formal academic text.
The asterisk: these scores are self-reported. humain-m3 has not been submitted to the Open Arabic LLM Leaderboard for independent evaluation. Verification only becomes possible after weights are released. That is a known pattern with new model announcements — but it is the honest context to carry into any evaluation.
Getting Access Now
The model is available through HUMAIN Node, which uses an OpenAI-compatible API — so existing code targeting the OpenAI SDK should work with minimal changes to the base URL and model ID.
Two access tiers exist in preview. The limited-preview tier applies Saudi alignment guardrails, disables thinking and streaming, and carries higher latency. The research-preview tier unlocks thinking modes, streaming, and lower latency. Most developers evaluating the model will want to request research-preview access. Neither tier has announced pricing.
HUMAIN Node also offers sovereign hosting — in-Kingdom data residency for Saudi and GCC enterprise customers who need it. For teams with strict data localization requirements, this closes a gap that ruling out ChatGPT or Anthropic APIs often creates.
Open Weights in October
HUMAIN targets an October 2026 open-weights release under the MiniMax Community License — contingent on completing safety training and alignment. Two things to know about that license: it is commercially restricted (not Apache 2.0 or MIT), and commercial use terms need to be read carefully before you build on top of it. “Open weights” here means you can download and run the model, not that you can use it freely for any commercial purpose.
The weights release matters for a second reason: it ends the benchmark debate. Once researchers can run humain-m3 independently, the self-reported 89.37% will either hold up or it will not. The October release is the real test.
The Honest Read
HUMAIN’s CEO said “this is not about importing a model and putting a Humain label on it.” At the architectural level, that is exactly what it is — humain-m3 is a fine-tuned derivative of a Chinese model, built with Saudi money. That is not a disqualifier. France’s Mistral, the UAE’s Falcon, and India’s Sarvam all exist on similar logic: national AI infrastructure built on a combination of local compute, local data, and global base models. The question is whether the Arabic post-training delivers real capability gains for Arabic-language tasks, and the benchmark numbers — self-reported as they are — suggest it does.
For developers building Arabic NLP products, the pragmatic call is this: request preview access, run your own evaluation on tasks that matter to your product, and watch the October weights drop. Do not confuse benchmark claims with production-readiness, and read the license before committing to a fine-tuning project on top of it.
Three Steps for Arabic NLP Teams
- Request research-preview access at node.humain.com — limited-preview tier is live now
- Benchmark on your actual workload: the ALRAGE and Arabic EXAMS gaps mean RAG pipelines and academic tasks need custom evaluation before you swap providers
- Mark October in your calendar — the open-weights release is when independent verification and fine-tuning plans become real; check the Open Arabic LLM Leaderboard for scores once weights land
