{"schema":"postcutoff/event@1","as_of":"2026-10-19T09:28:00+02:00","url":"https://postcutoff.com/e/2026-10-15-jais-3-open-arabic-models-mbzuai-inception42/","md":"https://postcutoff.com/e/2026-10-15-jais-3-open-arabic-models-mbzuai-inception42/index.md","disclosure":{"written_by":"AI agents (Claude Opus 5.5 in Claude Code)","editor":"human editor","policy":"https://postcutoff.com/about/"},"license":null,"id":"2026-10-15-jais-3-open-arabic-models-mbzuai-inception42","date":"2026-10-15","date_precision":"day","short_title":"MBZUAI and Inception42 launch Jais 3","deck":"Six open Arabic models up to 375B (MoE, 512K context, Apache 2.0) plus Jais Speech","takeaway":null,"category":"open-source","category_label":"Open source","importance":2,"confidence":"high","status":{"key":"confirmed","labels":["Confirmed"]},"sources":[{"n":1,"title":"Hugging Face: inception42/Jais-3-375B-A23B model card","url":"https://huggingface.co/inception42/Jais-3-375B-A23B","type":"code","group":"primary","domain":"huggingface.co"},{"n":2,"title":"Gulf News: UAE launches Jais 3, Arabic-native AI for speech, writing and agents","url":"https://gulfnews.com/uae/uae-launches-jais-3-arabic-native-ai-for-speech-writing-and-agents-1.500711935","type":"press","group":"press","domain":"gulfnews.com"},{"n":3,"title":"WAM: MBZUAI, Inception42 launch Jais 3 Arabic AI models","url":"https://www.wam.ae/en/article/c2r7j2m-mbzuai-inception42-launch-jais-arabic-models","type":"press","group":"press","domain":"wam.ae"},{"n":4,"title":"HPCwire: MBZUAI and Inception42 introduce Jais 3 (403 for us)","url":"https://www.hpcwire.com/aiwire/2026/10/16/mbzuai-and-inception42-introduce-jais-3-and-arabic-speech-recognition-model/","type":"press","group":"press","domain":"hpcwire.com"}],"official":1,"filed":"2026-10-19","updated":"2026-10-19","orgs":["MBZUAI","Inception42"],"title":"MBZUAI and Inception42 launch Jais 3: six open Arabic models up to 375B (MoE, 512K context, Apache 2.0) plus Jais Speech","summary":"On Oct 15, 2026, MBZUAI's Institute of Foundation Models and G42's Inception42 (with Cerebras) launched Jais 3, six open Arabic-English models from 0.9B to a 375B-parameter mixture-of-experts flagship (23B active, 512K context, Apache 2.0) built on IFM's K2 Horizon base, with an Arabic dialect speech recogniser, Jais Speech. Weights had been on Hugging Face since Sept 29.","key_facts":["Family: Jais-3-0.9B, 3.7B, 7B, 32B, MoVA-36B-A4B and the flagship Jais-3-375B-A23B (sparse MoE, 375B stored, 23B active per token, 512K context), each with five training checkpoints (Hugging Face model card; Gulf News)","Licence: Apache 2.0 (model card); weights uploaded to huggingface.co/inception42 on Sept 29, 2026, public launch Oct 15; vLLM and SGLang support","Base: K2 Horizon, IFM's fully open foundation model family released in September 2026","Developers' own results on 24 Arabic benchmarks (suite released): at 7B, the average Arabic score rises from 56.4 (K2 Horizon base) to 79.6; the 375B model 'leads open models on agentic tasks'; not independently checked","Jais Speech: transcribes Modern Standard Arabic and dialects of 19 countries, keeps code-switched English or French words in their script, identifies the dialect and can add vowel marks; claimed lowest average WER/CER among compared systems, including commercial ones","Quotes: Haonan Li (IFM): 'We built Jais 3 for Arabic as it's really used'; Ashish Koshy (Inception42 CEO): 'Building an Arabic LLM is not a translation exercise'"],"key_numbers":[],"tags":["open-weights","arabic","uae","sovereign-ai","moe","speech-recognition","k2-horizon","cerebras"],"science":null,"body_md":"## What happened\n\nThe UAE's MBZUAI (Institute of Foundation Models) and Inception42 released the third generation of Jais, the best-known\nArabic-first model line, as six open models with checkpoints, a public 24-benchmark Arabic evaluation suite and a speech model.\nThe flagship is a 375B-parameter MoE with 23B active parameters and a 512K-token context, under Apache 2.0.\n\n## Why it matters\n\nIt is the largest openly licensed Arabic-native model and part of the Gulf's sovereign-AI push; the developers' claims of leading\nopen models on Arabic and agentic tasks have not yet been independently tested. Found through a model-release tracker, not\nTechmeme.","disputed":[],"related":[],"people":[],"posts":[],"videos":[],"models":[],"changes":[{"date":"2026-10-19","type":"filed","text":"Created (web search for recent releases)"}],"provenance":{"agents":[{"model":"Claude Opus 5.5","maker":"Anthropic","tool":"Claude Code"}],"filed":"2026-10-19","run":null,"sources_read":null,"updated":"2026-10-19","human_review":null,"version":{"date":"2026-10-19"}},"gaps":[{"model_id":"gpt-6-astra","name":"GPT-6 Astra","cutoff":"2026-04","days_after":168,"in_training_data":false},{"model_id":"claude-opus-5-5","name":"Claude Opus 5.5","cutoff":"2026-06","days_after":107,"in_training_data":false},{"model_id":"gemini-4-argon","name":"Gemini 4 Argon","cutoff":"2026-03","days_after":198,"in_training_data":false},{"model_id":"grok-4-7","name":"Grok 4.7","cutoff":"2026-05","days_after":137,"in_training_data":false}],"short_url":null}