HUMAIN · 2026-09-03 · major
humain-m3 — a 428B Arabic model HUMAIN commissioned from MiniMax
humain-m3 is a 428B mixture-of-experts model with 23B active parameters, commissioned by Saudi Arabia's HUMAIN and built by MiniMax. It averages 89.37% across seven public Arabic benchmarks and is in limited preview.
Saudi Arabia's HUMAIN commissioned a 428B Arabic frontier model from MiniMax and opened it as a limited preview.
Key specs
| Parameters | 428B (23B active per token) |
|---|---|
| Average across 7 arabic benchmarks | 89.37% |
Quick facts
| Maker | HUMAIN (Saudi Arabia) |
|---|---|
| Built by | MiniMax, on the MiniMax-M3 lineage |
| Size | 428B mixture-of-experts, 23B active per token |
| Extra pretraining | 1T+ tokens of Arabic-native content |
| Availability | Limited preview on HUMAIN Node |
| Weights | Planned under the MiniMax Community License |
| Announced at | LEAP, Riyadh |
Benchmarks
| humain-m3 | 97.53% | |
|---|---|---|
| GPT-5.6 SOL | 93.42% | |
| Opus 5 | 91.38% | |
| MiniMax M3 reference | 90.6% |
| humain-m3 | 90.7% | |
|---|---|---|
| GPT-5.6 SOL | 88.82% | |
| Opus 5 | 88.77% | |
| MiniMax M3 reference | 81.8% |
| humain-m3 | 86.45% | |
|---|---|---|
| GPT-5.6 SOL | 81.54% | |
| Opus 5 | 83.06% | |
| MiniMax M3 reference | 75.31% |
What is it?
humain-m3 is a 428-billion-parameter mixture-of-experts model, 23B active per token, that HUMAIN announced on September 3, 2026 at LEAP in Riyadh. It is built on the MiniMax-M3 lineage and further pre-trained on more than a trillion tokens of Arabic-native content. Access is a limited preview through HUMAIN Node, not a public launch.
How does it work?
Instead of training from scratch, HUMAIN paid MiniMax to continue pre-training M3 on Arabic data, then use the preview period to check capability, safety and alignment across Arabic dialects with early users. On HUMAIN's own seven-benchmark comparison the model averages 89.37% and leads five of the seven, with its widest margin on AraTrust at 97.53% against 93.42% for GPT-5.6 SOL.
Why does it matter?
Arabic is spoken by hundreds of millions of people yet stays thinly covered by frontier models, and HUMAIN CEO Tareq Amin names that gap as the reason for the project. The other signal is where the weights came from: a Gulf state's national AI programme licensed a Chinese lab's model rather than training its own. HUMAIN says it plans to publish the weights under the MiniMax Community License once safety and alignment work is done.
Who is it for?
teams building Arabic-language products
Frequently asked questions
- Can I download the humain-m3 weights?
- Not yet. humain-m3 is in a limited preview on HUMAIN Node, with tiered access that starts guardrailed and opens to a research preview for approved users. HUMAIN says it expects to release the weights under the MiniMax Community License once safety training and alignment are complete, a step it currently targets for the following month.
- Which Arabic benchmarks does humain-m3 not win?
- Two of the seven. On ALRAGE, a retrieval-augmented generation test, humain-m3 scores 94.63% against 94.91% for GPT-5.6 SOL. On Translated MMLU it scores 93.20% against 93.78% for Opus 5. HUMAIN's table shows the model ahead on the other five, including AlGhafa, ArabicMMLU, Arabic EXAMS, MadinahQA and AraTrust.
- How much does the Arabic pretraining add over the base MiniMax M3?
- A large amount on every benchmark HUMAIN published. Against the MiniMax M3 reference checkpoint, humain-m3 gains about 11 points on AlGhafa (86.45% vs 75.31%), roughly 9 points on ArabicMMLU (90.70% vs 81.80%), and over 15 points on ALRAGE (94.63% vs 79.32%). The average across the seven tests rises from 80.34% to 89.37%.
- Why did a Saudi lab use a Chinese model instead of building one?
- HUMAIN commissioned the model from MiniMax and had it delivered rather than pre-training a 428B model itself, which shortens the path to a frontier-scale Arabic system. HUMAIN frames humain-m3 as the first model open to preview users on HUMAIN Node, a platform it says already hosts more than 100 frontier, proprietary and open-source models with broader access opening soon.
Try it
Request preview access at https://node.humain.com/