Tasks and inputs
Jais-3-MoVA-36B-A4B is listed for Text generation, Translation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Jais-3-MoVA-36B-A4B is the sparse Jais 3 model: 36B total, 4B active per token, 512K context, tuned for Arabic-English agents and tool use on one H100 node.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Thirty-six billion parameters stored, four billion doing the work on any given token. Jais-3-MoVA-36B-A4B is the sparse branch of IFM's Jais 3 family, and the reason to look at it is simple: roughly the quality of the dense 32B sibling at a fraction of the compute per token.
MBZUAI's Institute of Foundation Models and G42's Inception42 released it under Apache-2.0.
IFM calls the attention design Mixture-of-Values attention, MoVA, and stops there. It does not describe the mechanism or give an expert count. We will not guess. What you can see is its effect on serving: the validated recipe needs expert parallelism.
Both this model and the dense 32B fit one H100 node with two GPUs. The difference is what each token costs.
| MoVA-36B-A4B | Dense 32B | |
|---|---|---|
| Active parameters per token | 4B | 32B |
| Stored parameters | 36B (metadata: 37B) | 32B (metadata: 35B) |
| Parallelism | TP=2 plus EP=2 | TP=2 |
| Arabic IFBench (strict) | 92.0 | 88.7 |
| MCPMark | 69.3 | 66.1 |
| BrowseComp-Plus | 71.8 | 71.9 |
| Terminal Bench 2.1 | 67.3 | 63.7 |
| Arabic MMLU | 87.7 | 88.6 |
| AraGen overall | 68.6 | 71.0 |
Publisher-reported numbers. The sparse model edges ahead on tool and agent benchmarks and slips slightly on knowledge and open-ended Arabic generation. Against the wider field, [Qwen3.8-27B] remains stronger on coding agents in IFM's own comparison, scoring 79.8 on Terminal Bench.
--tensor-parallel-size 2 --enable-expert-parallel, plus the k2_horizon reasoning and tool-call parsers.--tp 2 --ep 2, FlashAttention-3, and a model-override argument setting xllm_source_router_gemm_partitions to 2.reasoning_effort: "high"; the recommended output budget is at least 32,768 tokens.Expert parallelism adds an operational wrinkle the dense model does not have. Check that your serving stack's version supports it before you schedule a migration.
High-throughput Arabic-English agents: support desks that call several tools per conversation, browsing assistants, MCP-heavy workflows. Its best scores are precisely there. For legal drafting or knowledge-heavy Q&A the dense sibling is marginally safer.
The family's launch is reported for 15 October 2026; these weights had been public since 29 September.
Run the same 200 production prompts through this model and the dense 32B on identical hardware. Record tokens per second at your target concurrency, cost per completed task, and pass rate on your tool calls. If the pass rates are within a point, the sparse model's throughput usually decides it. For broader comparisons try Gemma 4 26B A4B, another 4B-active model, or the compare tool.
Compare the model on the work you need to complete.
Jais-3-MoVA-36B-A4B is listed for Text generation, Translation. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 524.29K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under Apache-2.0. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Tencent
Tencent Hy4 Flash is the cost-focused companion to Hy4, keeping vision and agent skills at a far lower token price. Availability is still unsettled.
Alibaba
Qwen4 Max (qwen4-max) is Alibaba's first Qwen4-generation flagship: 1M-token context, text, image and video input, thinking on by default, $1.50/$4.50 per 1M.
Sarvam
Sarvam-105B-1711 is a refreshed open post-train of Sarvam's 105B MoE, tuned for tool calling, instruction following, agents and Indian languages.
IBM
Granite 4.1 8B LLMON Preview is IBM's model trained on block-structured LLMON prompts to resist indirect prompt injection in agents and RAG.
NVIDIA
Nemotron-Labs Sigma 8B is NVIDIA's open continuous diffusion language model, denoising 64-token blocks in parallel; English, GPU-only, 8K context.
Alibaba
Qwen4 Max Preview was Alibaba's five-day early-access build of its Qwen4 flagship, offered only through a Token Plan subscription before GA qwen4-max.