ANVE: the private AI office, on the record. See what that means

ANVE

Offline AI in your language · updated 2026-10-16

Can offline AI handle Arabic?
العربية for offices with Gulf clients: formal Arabic, dialects and Arabic mixed with English.

Short answer. Yes for formal Arabic, with care for Gulf dialect. Whisper’s paper counts 739 hours of Arabic training audio and reports 16.0% word error rate on read sentences (FLEURS, large-v2) but 53.8% on Common Voice 9. On Emirati speech in the Casablanca study, zero-shot Whisper large-v3 scored 62.3%, and 90.9% on Arabic with English words mixed in. For text, Llama 4 lists Arabic, Qwen3 lists eight Arabic varieties, and three open models led from the Gulf exist: Fanar-2 from Qatar, Jais 2 from the UAE and ALLaM from Saudi Arabia. An office serving Gulf clients should test its own calls. ANVE has not tested Arabic yet. Coverage and figures are as of October 2026.

01

What open models cover in Arabic

TaskModelsWhat the source says
Speech to text: WhisperWhisper large-v2, large-v3Arabic (ar) is one code in Whisper’s tokenizer. The paper counts 739 hours of Arabic speech with Arabic transcripts and 2,286 hours paired with English text. Large-v2 scored 16.0% word error rate on FLEURS and 53.8% on Common Voice 9, a spread that shows how much the test set matters. The large-v3 card warns of uneven accuracy across accents and dialects.
Speech to text: Gulf and other dialectsWhisper large-v3, Meta MMSCasablanca study, zero-shot test sets: Whisper large-v3 scored 62.3% word error rate on Emirati speech (52.9% after the paper’s text clean-up), 69.9% on Yemeni, 48.4% on Jordanian and 87.2% on Moroccan; the average over its eight dialects was 69.5%. Meta’s MMS scored 85.4% on Emirati. On Saudi television speech (SADA), the best result in a 2025 study was MMS 1B fine-tuned on that data with a language model: 40.9%.
Speech to text: Arabic mixed with EnglishWhisper large-v3Casablanca’s segments with English words in Latin letters inside Arabic speech: 90.9% word error rate with automatic language detection, 131.5% when told the language is English and 103.6% when told it is Arabic.
Chat, summaries, drafting: general modelsLlama 4, Qwen3, gpt-oss, GemmaLlama 4’s card lists Arabic; Llama 3.3’s does not. Qwen3 lists eight Arabic varieties (Standard, Najdi, Levantine, Egyptian, Moroccan, Mesopotamian, Ta’izzi-Adeni, Tunisian). gpt-oss reports Arabic MMMLU of 82.7% (120b) and 76.3% (20b) at high reasoning effort. Gemma’s cards say over 140 languages without naming them.
Chat, summaries, drafting: models led from the GulfFanar-2, Jais 2, ALLaMFanar-2-27B-Instruct (Qatar, Apache 2.0, built on Gemma 3) says it supports Modern Standard Arabic and Gulf, Levantine and Egyptian dialects. Its own card reports ArabicMMLU of 74.67% against 73.08% for Qwen3-32B and 72.21% for Gemma 3 27B, and dialectal-Arabic reading comprehension (Belebele) of 86.81% against 85.98% and 85.54%, with no breakdown by dialect; these are the vendor’s figures. Jais 2 (MBZUAI and Inception in the UAE, with Cerebras; Apache 2.0; 8B and 70B chat sizes; gated download) says it covers Modern Standard Arabic, regional dialects and Arabic-English mixing. ALLaM 7B (Saudi Arabia, listed as Apache 2.0) has a 4,096-token context.
TranslationNLLB-200FLORES-200 and NLLB-200 list ten Arabic entries covering nine varieties: Modern Standard Arabic (in Arabic script and romanised), Egyptian, Moroccan, Najdi, Mesopotamian, North and South Levantine, Tunisian and Ta’izzi-Adeni, but none labelled Gulf. NLLB-200’s card gives a CC-BY-NC licence, calls it a research model not released for production, and says its translations cannot be used as certified translations.
02

Two Arabics: formal writing and Gulf speech

An office serving Gulf clients, for example a law firm, a chartered accountancy, a trading house or a hospital that treats patients from the Gulf, meets two Arabics. Contracts and letters tend to be in formal Modern Standard Arabic; calls and voice notes are in dialect. The figures above behave that way: 16.0% on read sentences, 53.8% on Common Voice, 62.3% on Emirati speech. Fanar says it supports Gulf dialect and Jais 2 says it covers regional dialects, but the text read for this page gives no Gulf-specific accuracy figure for either.

Arabic is written right to left, and English names, numbers and invoice references inside Arabic text change direction. Check how your documents and exports show them. Both Arabic digits (٠ ١ ٢ ٣) and Western digits (0 1 2 3) appear in business papers, so check which your clients expect.

Arabic is not an Indian scheduled language. The 2011 Census counted 54,947 people in India with Arabic as their mother tongue, grouped in a family of its own. Where Gulf clients’ data may be processed is a question for your office’s legal advisers.

03

What to test before you rely on it

  • Record real calls or voice notes from Gulf clients, with their consent, in the dialect they use. Do not rely on news-reader Arabic as a stand-in.
  • Check names: Arabic names in Latin letters have several accepted spellings, so compare the transcript with the spellings in your contracts and visas.
  • Check amounts and dates: dirhams, riyals, dinars, and Hijri dates if your contracts use them.
  • Record sentences that mix English words into Arabic, which scored 90.9% in the study, and decide how you want them written.
  • For contracts and official papers, use a human certified translator. NLLB-200’s own card says its output is not a certified translation.
  • Open an Arabic export in your word processor and check the right-to-left layout, punctuation and numbers.
04

Where ANVE stands

ANVE’s meeting-minutes maker handles Hindi, English and Hinglish today. Arabic has not been tested on ANVE yet, so there is no ANVE result to report and nothing here promises support. ANVE is pre-release: no sealed unit exists and there are no customers yet.

Client data on ANVE sits in one database on the box. A cloud model is optional: it runs only after an administrator adds a provider, its answers are labelled, and each call is logged as local or cloud. A policy that blocks confidential data from cloud models is not on ANVE yet, as are full-disk encryption and secure boot in the appliance image. The security page has the details.

A demo call is free. If Arabic for Gulf clients decides your purchase, ask for your own recordings to be tried before you commit.

05

Questions people ask

Can Whisper transcribe Gulf Arabic?
It accepts Arabic as one language, but in the Casablanca study zero-shot Whisper large-v3 scored 62.3% word error rate on Emirati speech and 69.9% on Yemeni. Test your own client audio before relying on it.
Which offline Arabic chat models come from the Gulf?
Qatar’s Fanar-2-27B-Instruct (Apache 2.0, says it supports Gulf dialect), Jais 2 from MBZUAI and Inception in the UAE, in 8B and 70B chat sizes (Apache 2.0, gated download), and Saudi Arabia’s ALLaM 7B (listed as Apache 2.0, 4,096-token context).
Can offline AI translate Arabic contracts?
It can draft a translation, but NLLB-200 calls itself a research model, is licensed CC-BY-NC, and says its output is not a certified translation. Contracts need a human certified translator.
Does Llama support Arabic?
Llama 4’s model card lists Arabic among its 12 supported languages. Llama 3.3’s card lists eight languages and Arabic is not one of them. Qwen3’s launch post lists eight Arabic varieties.
06

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

07

Sources