Jais-3-7B

What it is

Jais-3-7B is the dense 7B member of the Jais 3 family, with a context window of 512,000 tokens.^[1] Dense means that every parameter is used for every token the model processes, in contrast with mixture-of-experts designs that route each token through only part of the network. The context window is the amount of text a single request can hold at once, including the prompt, documents, code and the model's own output; at 512,000 tokens it is large enough for book-length input or a sizeable code repository. In BF16 precision the model occupies about 18 GB, so it can run on a single GPU.^[2] It is intended for Arabic and English, and the model card lists other languages as out of scope.^[1]

Where it comes from

The model was announced on 15 October 2026 alongside the rest of the Jais 3 family.^[3] Its training is documented stage by stage: checkpoints from the continued pre-training, cooldown, FTPO and SFT stages are published as branches of the repository.^[1] What is released is therefore not only the finished model but also the intermediate states at the end of each of those stages. For researchers studying how Arabic capability builds up over a single training run, the branches make the process directly observable instead of something that has to be inferred from the final weights.

How reasoning is returned

When the model thinks before answering, the thinking is returned in a separate reasoning_content field and the answer in content.^[1] Keeping the two apart lets an application display, log or hide the intermediate reasoning without having to parse it out of the answer text. Reasoning effort is set per request through the chat template, and IFM recommends the high setting for all the results it reports.^[1] The model card also recommends allowing at least 32,768 output tokens when thinking mode is enabled, because the reasoning text is itself generated output and counts against the output budget.^[1]

Results

On IFM's Arabic average, Jais-3-7B scores 79.4, against 66.7 for the previous flagship, Jais 2 70B.^[2] On IFM's agentic average it scores 49.5, against 21.0 for Gemma 4 12B.^[2] It scores 70.2 on SWE-Bench Verified and 67.7 on BrowseComp-Plus.^[2] On AA-LCR, an Arabic long-context benchmark, it scores 65 against 49 for Gemma 4 12B.^[2] The comparisons put the model against its predecessor on Arabic averages and against Gemma 4 12B on agentic and long-context tasks.

Limits and position

IFM warns that the model can produce inaccurate content and should not carry high-stakes decisions without human oversight.^[1] Language coverage is deliberately narrow: Arabic and English are supported, and other languages are out of scope.^[1] The size that lets it run on one GPU also explains why it is released next to the larger members of the same family rather than in place of them.^[2][3]