Amegilla
B AI's Arabic-native model family: text is available in B products, direct model access and voice are in private preview, and vision plus broader agent capabilities are planned.
Built for Arabic, not adapted to it.
Amegilla is trained from the ground up on real Saudi, Gulf, Levantine, and Egyptian Arabic - so it understands dialect, accent, and context the way a native speaker does.
"Hi, I'm Amegilla. I think in Arabic first - its dialects, its rhythm, its manners - and I answer plainly. No theatrics, no filler. Just clear, honest help."
What Amegilla can do.
Arabic-native across regional dialects
Fuṣḥā plus supported Saudi, Gulf, Levantine, and Egyptian dialects, read and written around real use.
Text now, voice in preview
Amegilla supports text, with voice in private preview. Vision and broader multimodal capabilities are planned.
Long memory
32K tokens of context - whole documents, long conversations, and full briefs held in mind at once.
Grounded & honest
Says what it knows, flags what it doesn't, and shows its working when the answer needs it.
Amegilla Arabic benchmark results.
In B AI's reported private-preview evaluation, Amegilla averages 72.5 across the seven displayed Arabic benchmarks. Review the evidence and limitations below before comparing models.
| Model | Params | ACVA | MMLU (ar) | EXAMS | MMLU | ARC-C | HellaSwag | AGIEval | Average |
|---|---|---|---|---|---|---|---|---|---|
| Amegillaours | TBA | 82.5 | 66.38 | 57.7 | 85.5 | 61.1 | 83.0 | 71.4 | 72.51 |
| ALLaM-Instruct | 70B | 79.01 | 62.23 | 58.47 | 75.43 | 59.56 | 84.97 | 65.67 | 69.33 |
| ALLaM-Instruct | 34B | 81.0 | 60.2 | 58.66 | 71.24 | 60.15 | 81.25 | 52.47 | 66.42 |
| ALLaM-Instruct | 13B | 78.59 | 51.03 | 54.93 | 61.8 | 55.89 | 81.14 | 48.42 | 61.69 |
| ALLaM-Instruct | 7B | 80.33 | 49.6 | 52.7 | 58.31 | 51.45 | 75.2 | 47.09 | 59.24 |
| Jais-Chat-v3 | 30B | 70.05 | 30.15 | 51.21 | 57.57 | 51.02 | 78.91 | 36.78 | 53.67 |
| Jais-Chat | 13B | 70.68 | 41.43 | 46.93 | 49.46 | 46.84 | 77.6 | 31.45 | 52.06 |
The average is the arithmetic mean of the seven displayed scores. Higher is better. Amegilla values are B AI-reported private-preview results; parameter count is not yet published.
Scores and evaluation suite follow ALLaM: Large Language Models for Arabic and English — International Conference on Learning Representations (ICLR 2025).
Evidence, methodology, and limitations
This disclosure separates the external comparison source from B AI's own preview-model claims.
- The comparison suite, task labels, and baseline-model rows follow the cited ALLaM paper. Use that primary paper for its evaluation setup and baseline details.
- The Amegilla row is reported by B AI for a private-preview model. It has not been independently evaluated or verified.
- Treat the table as directional until B AI publishes the model version, prompts, runtime settings, raw outputs, and a reproducibility package. Preview results can change.
Three ways to put it to work.
The same model, shaped into the tools people actually use - one to chat with, one to delegate to, and one to build with.
Chat that speaks your dialect.
Chat, draft, summarise, translate, and research in fuṣḥā and supported Saudi, Gulf, Levantine, and Egyptian dialects.
- Drafts, summaries, and translation
- Connects to your files and tools
- Web now; iOS and Android planned
An agent that uses your computer.
B Coworker is planned to handle tasks across real apps on Windows and macOS, with sensitive actions requiring confirmation.
- Works across your real apps
- Runs on Windows and macOS
- Asks before anything sensitive
A coding agent that speaks Arabic.
Describe what you want in Arabic, in your terminal - B Code reads your repo, writes the change, runs it, and explains what it did. An agentic coding partner, Arabic-native.
- Reads and edits your codebase
- Runs commands and tests
- Explains changes in Arabic
# Amegilla · reading 14 files…
✓ created auth/login.html
✓ created auth/session.js
✓ wired routes → /login
# تمّ - صفحة تسجيل الدخول جاهزة مع تحقّق من الجلسة.
$
One model family. Capabilities at different stages. B Chat and B Code run on Amegilla; B Coworker is planned.
Arabic AI model
Amegilla: an Arabic AI model for Saudi and Gulf dialects
Direct answer
Amegilla is the Arabic-native AI model family behind B AI products. Text is available through B Chat and B Code; direct model access and voice are in private preview, while vision is planned.
When comparing Amegilla with candidates for the best Arabic AI model or best Saudi AI model, review the published benchmark methodology and test each model on your real dialect, domain, safety, privacy, latency, and deployment requirements.
Sources and evidence
- ALLaM: Large Language Models for Arabic and English (ICLR 2025) Primary source for the comparison suite and baseline scores shown on this page.
Frequently asked questions
What is the Amegilla Arabic AI model?
Amegilla is B AI’s Arabic-native model family for language, dialect, reasoning, and tool use. Text is available in B products; direct model access and voice are in private preview, while vision and broader multimodal capabilities are planned.
Is Amegilla the best Arabic AI model?
There is no single best model for every task. Use the published benchmarks as one signal, then test Amegilla and alternatives on your own Arabic dialects, tasks, quality targets, and deployment constraints.
Does Amegilla support Saudi Arabic?
Amegilla is designed around Arabic and regional dialect use cases, including Saudi and Gulf Arabic, rather than adapting an English-first experience after training.
Meet Amegilla inside B Chat.
Create your account, verify your email, and continue into The B Organization workspace.