Jais-3-375B-A23B

What it is

Jais-3-375B-A23B is a sparse mixture-of-experts (MoE) language model. It stores 375 billion parameters in total, but only 23 billion of them are used for each token it generates^[1]. That split is what makes a model of this size servable: capacity grows with the stored parameter count, while the computation spent on any single token stays at the level of a far smaller network, because only a subset of the experts fires for each token.

Its context window is 512,000 tokens^[1]. In practice that means one input can carry hundreds of thousands of tokens before anything has to be truncated — long documents, extended tool-call transcripts, or large codebases. Serving it is still heavy work: it needs an eight-GPU H200 node and about 758 GB of memory in BF16^[2]. The weights are released under Apache 2.0, and an FP8 build is published separately^[1], which gives users a second, lighter-weight option for deployment.

Where it comes from

The model card sits in a Hugging Face repository dated 29 September 2026, more than two weeks before the announcement^[1]. The Institute of Foundation Models (IFM), which publishes the Jais line, describes the model as built to compete with frontier closed models on Arabic while staying an open release^[3]. That framing sets the target clearly: the comparison class it is aimed at is proprietary frontier systems, and the differentiator it claims is Arabic capability rather than openness alone.

How it scores

On IFM's Arabic average, Jais-3-375B-A23B scores 84.0, against 82.0 for Qwen3.5-397B-A17B and 78.7 for MiniMax M3^[2]. On the agentic average it reaches 60.2^[2] — ahead of the open models it is compared with, but behind GPT-5.6 Sol at 75.6^[2]. The individual agentic numbers are 70.2 on SWE-Bench Verified, 72.6 on BrowseComp and 68.5 on MCPMark^[2].

One methodological detail matters for reading these figures. At 375B, the BrowseComp-Plus benchmark used for the smaller models in the comparison is replaced by BrowseComp, and each model is run with its own harness^[2]. Agentic scores depend heavily on the scaffolding around a model — the tools, prompts and loop that drive it — so a per-model harness makes the numbers a measure of the whole setup each vendor ships, not of the raw model in isolation.

Where it sits

Three things define the model's place. First, scale with sparse activation: 375 billion stored parameters and 23 billion per token put it in the large-MoE class, where a single model can be both broad and affordable to run per token. Second, a 512,000-token context, which moves long-document and long-trajectory agent work into the range a single request can hold. Third, an Arabic-first evaluation target set against closed frontier models, released openly under Apache 2.0 with a separate FP8 build^[1].

The agentic results sketch the trade-off. On Arabic the model is placed at the top of the comparison IFM published^[2]; on agentic tasks it beats the open models listed but remains clearly below the strongest closed system in that table^[2]. The gap of roughly fifteen points on the agentic average is the honest summary of where an open Arabic-centric release stood against the closed frontier at the time of publication.

The repository date is a small but telling detail: weights and model card were on Hugging Face more than two weeks before the announcement^[1], so the release arrived with the artifacts already in place for anyone watching the repository.