Jais 3 ships six open Arabic models and a speech companion

MBZUAI’s Institute of Foundation Models has released Jais 3, an Arabic-centric model family spanning 0.9B to 375B parameters. The Jais 3 notes list an Apache 2.0 license, intermediate checkpoints, training recipes, logs, and a cleaned Arabic benchmark suite.

Jais 3 targets chat, document processing, coding, browsing, and tool use. A companion speech-recognition model, Jais-ASR, adds transcription across Modern Standard Arabic and regional dialects. Developers can use the same model family across edge hardware, local workstations, enterprise servers, and an eight-GPU H200 node.

Six sizes, two architecture styles

All six models inherit their architecture from corresponding K2 Horizon checkpoints and use the same five-stage Arabic adaptation recipe.

Dense models activate every parameter for each token. Mixture-of-experts models route tokens through a subset of expert blocks, making active parameter count a rough guide to per-token compute. Total parameter count still determines weight storage. BF16 is a 16-bit format; inference also requires runtime memory and a key-value cache that grows with the prompt.

The 0.9B model supports a 128K-token context window. The other five support 512K tokens, enough for hundreds of pages of text or long agent runs with accumulated tool output. Available memory and attention cost determine how much of that nominal window a deployment can use.

Arabic scores rise, with gaps at the frontier

IFM’s released tables provide the figures in this section using its accompanying evaluation harness and cleaned datasets. On the audited Arabic aggregate, the 375B model scores 84.0, compared with 82.0 for Qwen3.5-397B. It also scores above Claude Sonnet 5, with GPT-5.6 Sol and Gemini 3.8 Flash remaining ahead.

Among large open models, the 375B variant ranks second on the reported agentic aggregate behind Qwen3.8-Flash-Next and leads the Arabic file-processing comparison. Agentic evaluations measure work such as operating a terminal, editing code, browsing the web, and processing files through tools.

The smaller variants post the largest gains within their comparison groups. Jais 3 3.7B leads Qwen3.5-4B by 13 points on Arabic and ranks first across the benchmarks in IFM’s comparison. The 0.9B model outscores every evaluated model in the 3B to 5B range. Jais 3 7B reaches 79.4 on Arabic, up from 66.7 for Jais 2 70B, and exceeds Gemma 4 31B and Muse Glimmer 30B on the reported agentic average.

Terminal-Bench 2.1 evaluates command-line work, SWE-Bench Verified measures fixes to real GitHub issues, and BrowseComp-Plus tests browser-based research. The 3.7B and 7B variants rank first on each agent benchmark in IFM’s same-size comparisons.

The 375B model’s reported factuality aggregate on Arabic SimpleQA and HalluScore is 51.0. Its AA-LCR score, which measures Arabic long-context reasoning, trails Gemini and GPT-5.6. At the 32B scale, Qwen3.8-27B remains stronger on general agentic evaluations such as SWE-Bench Verified.

Five stages retune K2 Horizon

Jais 3 is IFM’s first adaptation of the open K2 Horizon family to another language. Each model begins with the final post-trained K2 Horizon checkpoint at the corresponding size, then passes through five stages:

1. Continued pre-training: 201.45B tokens, comprising 151.09B Arabic tokens and 50.36B tokens replayed from K2 Horizon’s mid-training data. Ablation tests compared Arabic-to-replay ratios of 90/10, 75/25, and 50/50 before IFM selected 75/25.
2. Cool-down: A 3.47B-token pass over chat-formatted data, Wikipedia, and task-oriented question answering.
3. Supervised fine-tuning: 41.1B tokens across roughly 22 million examples covering dialects, Arabizi, iʿrāb grammatical analysis, poetry meter, long-document question answering, and agent trajectories.
4. FTPO preference training: A token-level method based on Liquid AI’s Antidoom recipe. FTPO applies its training signal where a repetition loop begins. Across the six models, the average loop rate falls from 13.8% to 1.2%.
5. GRPO reinforcement learning: Group Relative Policy Optimization with verifiable rewards for instruction following, iʿrāb, mathematics, diacritization, and code.

Dialect-focused sources account for about 36% of the Arabic training tokens, including Arabizi, parallel dialect data, and transcribed speech. Separately, synthetic or machine-translated material accounts for roughly 59%, expanding the available dialect and reasoning data beyond native web corpora.

Cleaning 679 benchmark tasks

IFM reviewed about 50 Arabic and 20 English benchmark groups, covering 679 tasks. Questions were filtered when three frontier models independently classified them as invalid. On Arabic MMLU, corrections to 765 of 14,455 questions changed one internal checkpoint’s score by 4.03%.

The evaluation harness uses generated-answer scoring across every benchmark. Log-probability scoring infers a multiple-choice selection from token likelihoods; generation scoring evaluates the answer the model emits. IFM released the cleaned benchmarks and evaluation code with the model family, allowing teams to reproduce the reported procedure.

Jais-ASR adds dialect-aware speech

The companion Jais-ASR notes describe an open speech-recognition model built on a Jais 3 backbone. It covers Modern Standard Arabic and dialects from 19 Arab countries, identifies dialects, adds optional diacritics, and handles code-switching with English, French, Spanish, and Italian.

Under the Open Universal Arabic ASR Leaderboard protocol, IFM reports a 22.85% word error rate and an 8.37% character error rate across six public test sets. Lower values indicate fewer transcription errors, and the macro average gives each test set equal weight. Those results place Jais-ASR ahead of ElevenLabs Scribe v2 and Speechmatics Melia 1 under the same protocol.

For silence, music, and other audio without transcribable speech, the model returns [IFM-NoTranscript]. That explicit output reduces the fabricated transcripts associated with Whisper-style speech systems.

The documented vLLM deployment path pins the serving packages and loads a Jais-ASR reasoning parser:

pip install 'vllm==0.23.0' 'transformers==5.5.4' accelerate soundfile soxr
pip install ./vllm_plugin
vllm serve IFM/Jais-ASR --trust-remote-code --reasoning-parser jais_asr_tag --max-model-len 2048

The ./vllm_plugin path must resolve to the plugin’s local directory before installation.

Checkpoints, code, and license terms

The intermediate checkpoints let researchers inspect each adaptation stage, resume training from a chosen point, and compare changes introduced by supervised tuning, preference optimization, or reinforcement learning. The five-stage recipe also provides a concrete reference for teams adapting K2 Horizon to other languages.

The cleaned evaluation suite supports internal comparisons on the same Arabic tasks used in the release. Smaller Jais 3 variants provide local options for tool-oriented applications, while the 32B, 36B-A4B, and 375B-A23B models cover progressively larger server deployments.

The weights, checkpoints, recipes, and training logs are available in the Hugging Face collection. Apache 2.0 permits commercial use, modification, and redistribution subject to its license and notice terms. IFM credits Inception and Cerebras Systems for infrastructure support.