Jais 3 MoVA 36B-A4B
Jais 3 MoVA combines a mixture-of-experts model with Mixture-of-Values attention. The developer lists 36B stored parameters and 4B active per token, with an advertised 512K context window.
What the release includes
The card supplies tensor-parallel and expert-parallel settings, plus separate reasoning and tool parsers. Its starred evaluation results use audited, cleaned benchmark sets; those numbers should not be mixed with unmodified-set scores.Jais 3 MoVA 36B-A4B official model card
Plan for stored weights and expert execution
Practical guidance. The active count describes computation per token. It does not mean this model has the storage needs of a dense 4B model. Use the developer's expert-parallel serving recipe as the starting point for capacity measurements.Jais 3 MoVA 36B-A4B official model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | Jais-3-MoVA-36B-A4B |
| Architecture | Mixture of experts with Mixture-of-Values attention; 36B total, 4B activeJais 3 MoVA 36B-A4B official model card |
| Published context | 512K; developer notationJais 3 MoVA 36B-A4B official model card |
| Weight license | Apache 2.0Jais 3 MoVA 36B-A4B official model card |
Related models
- Jais 3 32B
Inception42 (G42) · Jais 3
Largest dense Jais 3 model, with 32 billion parameters and a stated 512K context window.
Launch date not confirmedDownloadable - Jais 3 375B-A23B
Inception42 (G42) · Jais 3
Largest Jais 3 sparse model, with 375 billion stored parameters and 23 billion active per token.
Launch date not confirmedDownloadable