Jais-ASR
Jais-ASR is a speech-to-text model (5.7B total) from IFM, released under the Apache-2.0 license. The reviewed source lists Arabic (MSA + 19 dialects), plus English, French, Spanish, Italian (5 explicitly enumerated in our record), and the model is designed for batch-first transcription. It is a specialist rather than a general pick: its best placement in the common setups we test is #24 of 26, for subtitles with timing. Use the matcher with your own language and hardware to see where it fits.
Arabic specialist covering Modern Standard Arabic and dialects from 19 Arab countries, with code-switching into English, French, Spanish and Italian. Apache-2.0; plan on roughly 12 GB of GPU memory in BF16, with FP8 and 4-bit AWQ builds also published. Ratings are provisional until independent benchmarks are available.
Key facts
| Provider | IFM |
|---|---|
| Model family | Jais |
| Parameters | 5.7B total |
| License | Apache-2.0 · Permissive — license text ↗ |
| Language coverage | Arabic (MSA + 19 dialects), plus English, French, Spanish, Italian · 5 enumerated in our record |
| Streaming | Batch-first (no streaming path in the reviewed source) |
| Timestamps | Not documented |
| Speaker diarization | No — needs a separate component |
| Speech translation | No |
| Documented runtimes | Hugging Face Transformers 5.16.1, vLLM 0.23.0 + Jais-ASR plugin — how to run it |
| Primary source | huggingface.co/IFM/Jais-ASR-FP8 ↗ |
| Rating status | Provisional (added November 3, 2026) |
| Last reviewed |
Hardware and workload fit
These are Speech Model Match's conservative 1–5 editorial ratings, relative to the other 19 reviewed models. They come from model size, official memory guidance and documented runtimes — not from a universal benchmark. How the ratings work.
| Dimension | Rating | Scale |
|---|---|---|
| CPU-only fit | 1/5 | |
| Apple Silicon fit | 2/5 | |
| NVIDIA GPU fit | 4/5 | |
| Speed | 3/5 | |
| Accuracy (relative) | 3/5 | |
| Low memory use | 2/5 | |
| Long recordings | 3/5 | |
| Live / low latency | 1/5 | |
| Subtitle work | 2/5 | |
| Edge / low-power devices | 1/5 |
Where it is strong
- NVIDIA GPU fit: rated 4/5
What to check first
- No streaming path in the reviewed source — plan for file/batch transcription
- CPU-only fit: rated 1/5
- Apple Silicon fit: rated 2/5
- Low memory use: rated 2/5
- Live / low latency: rated 1/5
- Subtitle work: rated 2/5
- Edge / low-power devices: rated 1/5
How it ranks in common setups
Ranks come from the same deterministic matcher that runs on the homepage, applied to all 26 reviewed models. They are decision guidance, not accuracy claims.
| Setup | Rank | Fit score | Top match |
|---|---|---|---|
| Long recordings on an 8–12 GB NVIDIA GPU, balanced | #25 of 26 | 64/100 · Possible fit | Whisper turbo |
| Live captions or voice input (native streaming + timestamps, speed first) | #26 of 26 | 47/100 · Weak fit | Nemotron 3.5 ASR Streaming 0.6B |
| Subtitles with timing, accuracy first | #24 of 26 | 61/100 · Possible fit | Whisper medium |
| Meetings with speaker labels (diarization + timestamps) | #24 of 26 | 61/100 · Possible fit | VibeVoice-ASR 7B |
| CPU-only machine, low memory use | #25 of 26 | 47/100 · Weak fit | VibeVoice-ASR-BitNet |
| Apple Silicon with 8–16 GB unified memory, long recordings | #25 of 26 | 54/100 · Weak fit | Whisper small |
| Japanese live transcription (native streaming, speed first) | Not ranked (language not listed) | — | Nemotron 3.5 ASR Streaming 0.6B |
Supported languages
Compare coverage across every reviewed model in the language support matrix.
Community activity on Hugging Face
- Downloads, last 30 days
- 10,751
- Likes
- 22
- Repository last updated
- 13 Oct 2026
- Published on the Hub
- 6 Oct 2026
Figures from the Hugging Face Hub API, 11 Nov 2026. Source: huggingface.co/IFM/Jais-ASR ↗. Downloads count file downloads, not active users, and do not affect the matcher's scores. Compare all models.
How to run Jais-ASR locally
IFM documents Transformers and a vLLM plugin that exposes an OpenAI-compatible transcription endpoint. Runtimes below are the ones the official model card or repository documents for this model (checked November 2, 2026).
- Hugging Face Transformers 5.16.1 ↗Framework · pinned version on the card
- vLLM 0.23.0 + Jais-ASR plugin ↗Serving engine · /v1/audio/transcriptions endpoint
Install command shown in the official huggingface.co page ↗:
pip install 'transformers==5.16.1' accelerate soundfile soxr
Audio input: Language settings ar, msa, en, fr, es, it plus 19 Arabic dialect codes; built to transcribe rather than translate.
Official hardware guidance: Plan on roughly 12 GB of GPU memory in BF16; FP8 (7.68 GB) and 4-bit AWQ (5.85 GB) downloads are also published. Source ↗
Whatever runtime you choose, measure speed and accuracy on your own audio with the benchmark tool; results differ a lot between runtimes, precision and hardware.
Before you deploy Jais-ASR
A printable checklist built from the facts on this page. Tick it off before putting the model into production.
- Language. Confirm your language is in the 5-language list on this page, and test any accents or code-switching you expect.
- License. Apache-2.0: Give recipients a copy of the license, keep copyright and attribution notices, carry forward any NOTICE file and mark files you changed. The patent license ends if you start patent litigation over the work. License guide.
- Install. Start from the official command:
pip install 'transformers==5.16.1' accelerate soundfile soxr(full run notes). - Audio input. Language settings ar, msa, en, fr, es, it plus 19 Arabic dialect codes; built to transcribe rather than translate.
- Hardware. Plan on roughly 12 GB of GPU memory in BF16; FP8 (7.68 GB) and 4-bit AWQ (5.85 GB) downloads are also published.
- Features you rely on. For Jais-ASR, no streaming path is documented; timestamps are not documented; speaker labels need a separate diarization component. Test each one you need with real audio.
- Measure. Run 10–20 representative clips, score them with the WER calculator and record speed in the benchmark workspace before you commit.
- Provisional ratings. This model was added on November 3, 2026; treat its hardware and workload ratings as estimates and weight your own measurements higher.
Alternatives and related models
Similar fit profile: Granite Speech 5.0 2B (IBM, 2B), Cohere Transcribe 03-2026 (Cohere Labs, 2B), VibeVoice-ASR 7B (Microsoft, 9B total). Similarity is computed from the ten fit ratings plus streaming and diarization support.
| Fact | Jais-ASR | Granite Speech 5.0 2B | Cohere Transcribe 03-2026 |
|---|---|---|---|
| Provider · size | IFM · 5.7B total | IBM · 2B | Cohere Labs · 2B |
| License | Apache-2.0 Permissive | Apache-2.0 Permissive | Apache-2.0 Permissive |
| Languages | Arabic (MSA + 19 dialects), plus English, French, Spanish, Italian | 6 listed languages | 14 listed languages |
| Streaming | Batch-first | Batch-first | Batch-first |
| Timestamps | Not documented | Not documented | Limited |
| Speaker diarization | Separate pipeline | Separate pipeline | Separate pipeline |
| CPU · Apple · NVIDIA fit | 1/5 · 2/5 · 4/5 | 2/5 · 4/5 · 5/5 | 1/5 · 2/5 · 5/5 |
| Live · long recordings | 1/5 · 3/5 | 1/5 · 3/5 | 2/5 · 5/5 |
| Main runtime | Hugging Face Transformers 5.16.1 | Hugging Face Transformers ≥ 4.52.1 | Hugging Face Transformers ≥ 5.4 |
Check Jais-ASR for your setup
Pick your language, hardware and workload to see this model's fit score, the reasons behind it and where it ranks against all reviewed models. Everything runs in your browser.
Common questions
Can I use Jais-ASR commercially?
Permissive license; commercial use is allowed under the listed terms. The listed license is Apache-2.0; read the current license text ↗ before shipping a product, because terms can change. The license guide compares the obligations of all 26 reviewed models.
Does Jais-ASR support real-time streaming?
No streaming path is documented in the reviewed source. Treat it as a file/batch transcription model.
Which languages does Jais-ASR transcribe?
Arabic (MSA + 19 dialects), plus English, French, Spanish, Italian. The full list of 5 explicitly enumerated languages is shown above.
Can Jais-ASR tell speakers apart?
Not natively. Combine it with a dedicated diarization component and evaluate that component separately; the meetings guide covers this workflow.
Primary sources
Reviewed against primary sources on November 3, 2026. Model behavior, runtime support, language coverage and licenses can change; re-check the linked primary source before production use. See the full source directory.