2026-10-15 to 2026-10-15 · Published 2026-10-16

Jais 3 added to BenchGecko · six open Arabic-centric models from 0.9B to 375B

MBZUAI Institute of Foundation Models released Jais 3 on 2026-10-15: six open-weight Arabic-centric models under Apache 2.0, up to 375B total and 23B active parameters, with 512K context on five of the six.

What launched

MBZUAI's Institute of Foundation Models released Jais 3 on 2026-10-15, with Inception as partner. It is a family of six open-weight Arabic-centric models built on IFM's K2 Horizon models, trained to work as agents in Modern Standard Arabic, regional dialects and Arabizi. Final weights, checkpoints from every training stage, configs and evaluation code ship under Apache 2.0. IFM publishes no hosted API price, so BenchGecko leaves the pricing fields empty.

The six models on BenchGecko

Jais 3 375B-A23B · MoE, 375B total, 23B active · about 758 GB in BF16 on one node of 8 H200 GPUs
Jais 3 36B-A4B · MoE layers with MoVA attention, 36B total, about 4B active · about 75 GB in BF16
Jais 3 32B · dense · about 70 GB in BF16
Jais 3 7B · dense · 18 GB in BF16
Jais 3 3.7B · dense · 10.1 GB in BF16
Jais 3 0.9B · dense · 2.2 GB in BF16

Context is 512K tokens for five models, stored as 524,288 on BenchGecko. The 0.9B has 128K, stored as 131,072. All six take text in and return text.

Vendor-reported results

IFM ran every comparison model with its own harness. BenchGecko records these as vendor claims and has no Gecko Tests for Jais 3 yet.

Arabic average, flagship tier:

Jais 3 375B-A23B · 84.0
Qwen3.5-397B-A17B · 82.0
Claude Sonnet 5 · 82.4
GPT-5.6 Sol · 88.2
Gemini 3.8 Flash · 89.1

Agentic results for the 375B: agentic average 60.2, Terminal-Bench 2.1 at 73.8, SWE-Bench Verified at 70.2, BrowseComp at 72.6 and GDPval at 73.2. Among the large open models in the table, only Qwen3.8-Flash-Next has a higher agentic average, at 68.8.

At smaller sizes, Jais 3 32B scores 83.1 on the Arabic average against 82.6 for Gemma 4 31B, and Jais 3 7B scores 79.4 against 66.7 for Jais 2 70B.

Training

Each model starts from the post-trained K2 Horizon model of the same size. Continued pre-training used 201.45B tokens, 75% of them Arabic. A final token preference step aimed at repetition loops cut the loop rate from 13.8% to 1.2% averaged over the six models, before reinforcement learning with verifiable rewards.

Sources

Source: BenchGecko

Written automatically from BenchGecko's tracked data for this period; every number comes from the changelog, model, mindshare and pricing datasets. Checked against the source data before publishing.