IFM and Inception42 launch Jais 3: six open Arabic-native agentic models (0.9B–375B) plus Jais Speech, 512K context on Hugging Face

Background

Most frontier models are English-first and only later patched for Arabic, which is brittle on dialects, cultural knowledge, and real office files. For more than 400 million Arabic speakers—and for hospitals, companies, and governments that need on-prem or sovereign stacks—builders want open models that can hear dialects, write constrained documents, and finish multi-step tool work, not chatty translation layers.

Architecture

Jais 3 is the first language adaptation of IFM’s open K2 Horizon fleet. Six sizes share one training recipe and agent stack: 0.9B–32B dense, 36B-A4B with MoE plus MoVA attention, and flagship 375B-A23B sparse MoE (~23B active). Context is up to 512K tokens (128K on 0.9B). Training runs CPT → cooldown → SFT → FTPO (against doom loops) → verifiable-reward RL. Agentic skill mainly comes from ~425K real tool trajectories in SFT, not from the RL stage.

Benchmarks

Scores are IFM same-harness runs on audited Arabic suites—treat them as vendor numbers until third parties reproduce them. Official Arabic averages: 375B leads open peers at 84.0 vs Qwen3.5-397B 82.0 and sits near some closed frontier models; agentic average 60.2 with SWE-Bench Verified 70.2 and Terminal-Bench 2.1 73.8. The 7B reaches 79.4 Arabic average, up from 56.4 on the same-size K2 Horizon base. Jais Speech posts macro-average WER/CER 22.85/8.37 across six public ASR sets. No independent leaderboard replay was available at write time.

Getting started

Weights and stage checkpoints ship Apache-2.0 on Hugging Face collection IFM/jais-3 (e.g. inception42/Jais-3-375B-A23B). vLLM (≥0.30) and SGLang (≥0.5.20) publish day-0 recipes; enable the k2_horizon reasoning and tool-call parsers and prefer reasoning_effort=high. Small models fit single-GPU BF16; the flagship’s reference serve is 16×H100 or one 8×H200 node. Pair speech with github.com/ifm-ai/JaisVoice. Details: ifm.ai/blog/jais3 and ifm.ai/blog/jais-asr.