Language modelLimited research preview
HUMAIN M3
HUMAIN
- Released
- September 3, 2026
- Data date
- October 7, 2026
- Input
- Text, image, and video
- Output
- Text in the published playground example
HUMAIN M3 is a research preview focused on Arabic, commissioned by HUMAIN from MiniMax. It handles text, images, and video, uses a MoE architecture with 428 billion total and 23 billion active parameters, and received further pretraining on more than one trillion Arabic-native tokens.
Access currently runs through the HUMAIN Node playground and its OpenAI-compatible API. The weights are not public yet; HUMAIN says a later release is planned under the MiniMax Community License.
Specifications and access
| Specification | Value and source |
|---|---|
| API model ID | humain-m3Source |
| Development | Commissioned by HUMAIN and delivered by MiniMax; built on the MiniMax M3 lineage.Source |
| Parameters | 428B total parametersSource |
| Active parameters | 23B per tokenSource |
| Context window | Not publishedSource |
| Input modalities | Text, image, and videoSource |
| Output modalities | Text in the published playground exampleSource |
| Training stage | Further pretraining on more than 1T Arabic-native tokens and Arabic post-training.Source |
| Weights | Not public yet; HUMAIN says it plans to release weights after the preview.Source |
| Announced license | MiniMax Community License is planned for the announced weights.Source |
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
AlGhafa (Show measurement, test conditions, and source)
- Source value
- 86.45
- Score
- 86.45
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
ArabicMMLU (Show measurement, test conditions, and source)
- Source value
- 90.7
- Score
- 90.7
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
Arabic EXAMS (Show measurement, test conditions, and source)
- Source value
- 67.67
- Score
- 67.67
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
MadinahQA (Show measurement, test conditions, and source)
- Source value
- 95.44
- Score
- 95.44
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
AraTrust (Show measurement, test conditions, and source)
- Source value
- 97.53
- Score
- 97.53
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
ALRAGE (Show measurement, test conditions, and source)
- Source value
- 94.63
- Score
- 94.63
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
Translated MMLU (Show measurement, test conditions, and source)
- Source value
- 93.2
- Score
- 93.2
- Metric
- published score
- Unit
- %
- Category
- source-specific
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- HUMAIN Node official model and evaluation page
- Status
- active
- Retrieved at
- 2026-10-07
- Methodology
- Vendor-reported result. HUMAIN attributes this individual result to its evaluation of the previewed humain-m3 checkpoint. The provider lists it as one of seven public Arabic benchmarks in an equally weighted suite. No benchmark-specific prompt, decoding settings, repeat count, or sample size is published.
- Context
- Vendor-reported preview result. It is not an independent reproduction, and the page does not disclose the per-benchmark run protocol or sample size.
Sources and data date
| Type | Evidence and data date |
|---|---|
| Input | Documented modalities. Retrieved October 7, 2026. |
| Output | Documented modalities. Retrieved October 7, 2026. |
| Research status | Research date October 7, 2026. 4 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | per-benchmark-run-configuration-and-sample-counts Original wording HUMAIN identifies the previewed humain-m3 checkpoint and publishes seven public Arabic benchmark names plus percentages, but gives no benchmark-specific prompts, decoding settings, repeats, sample counts, or confidence intervals. Its equal-weighted suite average is reported by the provider but is not entered as a separate benchmark. |
| Unresolved | independent-exact-checkpoint-results Original wording An independent report evaluates the alias humain-m3-preview on September 5, 2026, but does not bind it to a revision of the canonical humain-m3 profile. Najd Arena explicitly labels its results historical pre-audit evidence with unverified evaluation dates and case contents that differ from the current dataset. Neither result was transferred. |
| Additional source | HUMAIN announcement issued through PR Newswire (retrieved October 7, 2026) · HUMAIN announcement issued through PR Newswire · Editorial description reviewed October 7, 2026 |