Jais-ASR
IFM's Apache 2.0 speech recognizer for Arabic as it is spoken: dialects from 19 countries, code-switching and optional diacritics.
In short. Jais-ASR turns spoken Arabic into text, including Modern Standard Arabic and dialects from 19 Arab countries, even when speakers slip into English, French, Spanish or Italian. It can add vowel marks, tell you which language or dialect it heard, and says so when a clip has no speech instead of inventing words.
Summary
Jais-ASR is a speech recognition model from the Institute of Foundation Models (IFM) at MBZUAI. It was announced alongside Jais 3 on 15 October 2026, and its Hugging Face repository first appeared on 6 October. It is based on Jais-3-4B, the 3.7B Jais 3 language model, and the full model, including its audio components, has about 5.7 billion parameters and takes 11.4 GB in BF16. The card describes Modern Standard Arabic and dialects from 19 Arab countries, with natural code-switching into English, French, Spanish and Italian. Optional switches add diacritics and keep foreign-language spans in their own script. Each transcript comes with a language or dialect label. When a clip holds only music or noise, the model returns a no-transcript marker and a refusal flag instead of made-up text.
IFM reports 22.85% average word error and 8.37% average character error on the six test sets of the Open Universal Arabic ASR Leaderboard, scored with the leaderboard's own script. In IFM's table that is ahead of Audar-ASR-V1-Turbo at 23.17%, Speechmatics Melia 1 at 23.51% and ElevenLabs Scribe v2 at 24.54%, then a cohere-transcribe-arabic build at 25.87%, OmniASR-LLM-7B at 28.32%, Qwen3-Omni-30B-A3B at 30.71%, Qwen3-ASR-1.7B at 33.36% and Voxtral-Small-24B at 34.47%. IFM measured Speechmatics and ElevenLabs itself and took the open models from the leaderboard. The per-set numbers are more mixed: on Common Voice 18 read speech Jais-ASR scores 8.56%, behind the Cohere build at 5.82% and Speechmatics at 5.19%, and on the Casablanca dialect set it scores 46.56% against 46.41% for Speechmatics.
You can run it with Transformers using trust_remote_code, or serve it through the included vLLM plugin behind an OpenAI-compatible transcription endpoint. The BF16 weights need roughly 12 GB of GPU memory. IFM also publishes an AWQ build at 5.8 GB, FP8 and MLX builds, and GGUF files from 2.9 GB (IQ4_XS) to 5.4 GB (Q8_0) that need a separate multimodal projector file. In the 30 days to 11 November 2026, Hugging Face counted about 10,800 downloads of the main model, 8,000 of the AWQ build and 3,000 of the GGUF files. In IFM's demo, Jais-ASR hands transcribed speech to Jais 3, which carries out the spoken instruction as an agent.
Good for
- Transcribing Arabic calls, broadcasts, lectures and interviews, including dialect-heavy and noisy audio
- Bilingual speech where Arabic mixes with English, French, Spanish or Italian
- Arabic text that needs vowel marks, or pipelines that want a dialect label with each transcript
- Voice-driven Arabic agents, paired with Jais 3
Watch out for
- The headline numbers are IFM's own, and Speechmatics and ElevenLabs were measured by IFM rather than taken from the leaderboard
- On clean read speech such as Common Voice it trails the best systems in IFM's table, so test on your own audio
- Real-time streaming is not described on the card, and at about 12 GB of GPU memory in BF16 it is much heavier than small recognizers such as Whisper large-v3-turbo
- It covers Arabic and four other languages, not general multilingual transcription
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Jais-ASR. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Jais-ASR (BF16) BF16 | 5.72B | 11 GB | 13 GB | Laptop or phone |
| Jais-ASR W4A16 AWQ INT4 | 5.72B | 5.8 GB | 6.67 GB | Laptop or phone |
| Jais-ASR GGUF Q4_K_M INT4 | 5.72B | 3.16 GB | 3.63 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is Jais-ASR?
- Jais-ASR turns spoken Arabic into text, including Modern Standard Arabic and dialects from 19 Arab countries, even when speakers slip into English, French, Spanish or Italian. It can add vowel marks, tell you which language or dialect it heard, and says so when a clip has no speech instead of inventing words.
- Can I use Jais-ASR commercially?
- Yes. Jais-ASR is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Jais-ASR open source?
- Jais-ASR is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Jais-ASR on my own computer?
- Yes. The smallest variant, Jais-ASR GGUF Q4_K_M, needs about 3.63 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Jais-ASR released?
- Jais-ASR was released by Institute of Foundation Models (MBZUAI) on October 15, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Jais-ASR model card (Hugging Face) huggingface.co
- Jais 3: Introducing Arabic-centric Agentic AI (IFM blog, 15 Oct 2026) ifm.ai
- Open Universal Arabic ASR Leaderboard (GitHub) github.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Jais 3 Institute of Foundation Models (MBZUAI) | 1.08B – 379B 23B active | 512K | Apache 2.0 (Permissive) | 0.745 GB Laptop or phone | |
| Cohere Transcribe Cohere | 2.07B | – | Apache 2.0 (Permissive) | 4.75 GB Laptop or phone | |
| Qwen3-ASR Alibaba Qwen | 918M – 2.35B | – | Apache 2.0 (Permissive) | 2.12 GB Laptop or phone | |
| Whisper large-v3-turbo OpenAI | 810M – 1.54B | – | MIT (Permissive) | 6 GB Laptop or phone |