Jais 3
modelYour notes
The third Jais generation, released 15 October 2026 by MBZUAI's Institute of Foundation Models with Inception42 (the model card's citation credits "IFM Team, Inception42 Team, and Cerebras Team"), and the first that is not its own pretraining run. Earlier generations (Jais, Jais 2) were trained from scratch. Jais 3 is Arabic-centric continued pretraining of K2 Horizon, IFM's six-model open fleet from six weeks earlier. Launch coverage says it is "built on the open-source K2 Horizon Fleet". The card names no base, but the config.json of the 375B flagship matches K2-Horizon-375B-A23B in every architectural field (K2HorizonForCausalLM, 61 layers, hidden size 6,144, 192 experts with 8 active, 250,624 vocabulary, 524,288 positions). Only initializer_range and the router auxiliary-loss coefficient differ. The size ladder is K2 Horizon's too: 0.9B, 3.7B, 7B, 32B, MoVA 36B-A4B and 375B-A23B, plus diffusion-augmented Uno variants of the four dense sizes. The parameter counts therefore stay on the K2 Horizon entry.
The flagship Jais-3-375B-A23B "stores 375B parameters and runs 23B per token, with a 512K context window", is a reasoning model (high effort recommended for all reported results), and ships under Apache 2.0 with BF16, FP8, GGUF and MLX builds. In keeping with IFM's 360-open practice, the repository exposes the CPT, Cooldown, SFT and FTPO checkpoints as branches, and the training logs are public on W&B. CPT token counts, data mix and compute are not stated on the card. The technical report was marked "coming soon" at launch; on 25 October the model card added a link to it at ifm.ai/jais/tech-report (cited as "Jais 3: Introducing Arabic-centric Agentic AI", IFM, Inception42 and Cerebras teams). The report reached arXiv on 6 November (2611.10224). It describes Jais 3 as six models from 0.9B to 375B, each adapted from the matching K2 Horizon checkpoint rather than pretrained from scratch. Each goes through Arabic continual pretraining, SFT, FTPO and RLVR. The data mixes natural, translated, transliterated and synthetic Arabic, and replays K2 Horizon's own training data to keep its reasoning, agentic and English skills. All models except the 0.9B keep the 512K context. The report calls the 375B the largest open Arabic-centric model to date. Evaluation moves to fully generative scoring on audited Arabic benchmarks, which are released with the evaluation code, data recipes, configs and logs. The code repository is still marked "coming soon". A subset of the Arabic training data is released as IFM/Jais-3-Data.
The card's numbers are self-reported against Qwen3.5-397B-A17B, MiniMax M3, Qwen3.8-Flash-Next and Nemotron 3 Ultra. On Arabic the model is roughly at parity rather than ahead: Arabic MMLU 89.3 (Qwen3.5-397B 89.1), Belebele 95.6, AraGen v3 72.4 (Qwen3.5-397B 78.9), Arabic GSM8K 86.8 (others 93–94). Its clear leads are dialect identification (QADI 71.1 against 62.3 at best), FLORES200+ translation 91.9, and IFM's own Arabic IFBench 93.9, which is not yet public. On agentic tasks it slightly improves on its K2 Horizon base: Terminal-Bench 2.1 73.8 (K2 Horizon 70.2), MCPMark 68.5 (67.7), BrowseComp 72.6 (72.8). It trails the 180B Qwen3.8-Flash-Next on most agent rows and SWE-Bench Verified (70.2 against 80.8). Many of the Arabic benchmarks are scored on audited subsets: IFM released more than a dozen "_verified" re-releases of existing Arabic evals with ambiguous or wrongly keyed items removed. That makes the numbers more defensible, but they are not directly comparable to scores on the original sets. The release also included Jais Speech (IFM/Jais-ASR), an Arabic speech recognition model. Not on Artificial Analysis at filing.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| Arabic MMLU (audited) | 89.3 | — |
| Belebele (Arabic) | 95.6 | — |
| AraGen v3 (3C3H) | 72.4 | — |
| Arabic IFEval | 77.7 | — |
| AA-LCR (Arabic) | 77.0 | — |
| Terminal-Bench 2.1 | 73.8 | avg@3, high reasoning effort |
| SWE-Bench Verified (offline) | 70.2 | — |
| MCPMark | 68.5 | — |
| BrowseComp | 72.6 | discard-all at 95K |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Jais-3-375B-A23B | — | flagship MoE, 375B stored / 23B per token, 512K context; CPT of K2-Horizon-375B-A23B; BF16 and FP8 |
| Jais-3-MoVA-36B-A4B | — | MoE with attention-value experts (K2 Horizon MoVA shape); FP8, GGUF, MLX |
| Jais-3-32B | — | dense; Uno, FP8, GGUF, MLX builds |
| Jais-3-7B | — | dense; card says 7B, Hugging Face safetensors metadata reports about 9B; Uno, FP8, GGUF, MLX builds |
| Jais-3-3.7B | — | dense; Uno, FP8, GGUF, MLX builds |
| Jais-3-0.9B | — | dense; Uno, GGUF, MLX builds |