Fanar-1-9B-Instruct Model Overview

Updated 9 July 2026

Fanar-1-9B-Instruct is a first-generation 9B Arabic-centric instruct model derived from Gemma-2-9B, notable for its continual pre-training and dual-language capabilities.
The model employs a decoder-only Transformer with vocabulary pruning and robust architecture details, delivering strong performance across Arabic knowledge and English world benchmarks.
Instruction tuning through SFT and DPO, supported by millions of curated samples, enables the model to excel in Arabic cultural, dialectal, and language tasks.

Fanar-1-9B-Instruct is the first-generation 9B instruct-tuned model in the Fanar family of Arabic-centric generative models. In the Fanar 2.0 paper it is identified explicitly as “the first generation Fanar model, built on top of gemma-2-9b base model using continual pre-training and instruction fine-tuning,” whereas the original Fanar platform paper uses the name “Fanar Prime Instruct” for the corresponding 9B-class aligned model rather than the exact string “Fanar-1-9B-Instruct” (Team et al., 17 Mar 2026, Team et al., 18 Jan 2025). This suggests that “Fanar-1-9B-Instruct” is the checkpoint-style name used in later benchmarking, while “Fanar Prime Instruct” is the earlier family/platform name for the same 9B instruct line.

1. Position within the Fanar lineage

Within the Fanar family, Fanar-1-9B-Instruct occupies the role of the stronger first-generation 9B model. The Fanar 2.0 paper places it inside Fanar 1.0’s dual-LLM strategy, consisting of Fanar Star (7B, trained from scratch) and Fanar Prime (9B, continual on Gemma-2-9B), before the architectural shift in Fanar 2.0 to a single 27B model (Team et al., 17 Mar 2026).

In that later comparison framework, Fanar-1-9B-Instruct is not treated as a minor ablation or auxiliary checkpoint. It serves as the principal first-generation baseline against which Fanar-27B is measured across Arabic knowledge, Arabic language, dialectal Arabic, Arabic and Islamic culture, English world knowledge, and conversational instruction-following tasks (Team et al., 17 Mar 2026). This historical placement matters because it fixes the model’s significance: it is the reference point for the transition from Fanar 1.0’s smaller Arabic-centric Gemma-2-derived line to Fanar 2.0’s larger Gemma-3-derived line.

The naming discrepancy between Fanar Prime Instruct and Fanar-1-9B-Instruct is central to the literature. The original platform paper does not use the latter string, but it does identify Fanar Prime Instruct as the 9B-class aligned model in the same family (Team et al., 18 Jan 2025). A plausible implication is that later benchmark-oriented naming standardized the model as Fanar-1-9B-Instruct, while the original system paper used the internal family names Star and Prime.

2. Architecture, initialization, and pretraining basis

Under the earlier Fanar platform nomenclature, the most detailed technical specification is given for Fanar Prime, the 9B-class model whose aligned form is Fanar Prime Instruct. Fanar Prime comprises 8.78B parameters total, specifically 459M embedding parameters, 8.32B non-embedding parameters, and 8.78B total parameters. It is built through the continual pre-training of the Gemma-2-9B-Base model rather than from-scratch training (Team et al., 18 Jan 2025).

The architecture is described as a decoder-only Transformer. For the 9B model, the paper reports model dimension 3584, RMSNorm for both pre- and post-normalization, RoPE, GQA, Approximated GeGLU, no biases, vocabulary size 128,256, and weight tying: False. A major initialization step is vocabulary pruning: the original 250,000-token Gemma vocabulary is reduced to 128,256 tokens, which in turn reduces total parameters from 9.2 billion to 8.78 billion (Team et al., 18 Jan 2025). By contrast, several architectural items remain unspecified or unreadable in the extracted description, including the exact number of layers, attention-head count, intermediate size, and the exact context-length entry for Fanar Prime (Team et al., 18 Jan 2025).

The pretraining basis is a cleaned and deduplicated multilingual corpus of 1 trillion tokens spanning Arabic, English, and code. The platform paper reports the overall corpus as roughly 50% English, 40% Arabic, and 10% code, and states that continual pretraining for Fanar Prime uses a similar balanced data composition of 45% Arabic, 45% English, and 10% code (Team et al., 18 Jan 2025). The Fanar 2.0 retrospective summarizes Fanar 1.0 more coarsely as ~1T tokens (40% Ar / 50% En / 10% Code) (Team et al., 17 Mar 2026). The papers therefore report the data mixture at different levels of granularity.

Arabic coverage in the corpus is broad by design. The Arabic portion is said to cover MSA, Classical Arabic, and dialectal Arabic, and to draw from web documents, news articles, books on literature, religion, politics, culture, and history, Arabic Wikipedia, classical and contemporary poetry, and machine-translated books, STEM papers, and encyclopedic documents (Team et al., 18 Jan 2025). The platform paper also emphasizes translation-augmented Arabic coverage, including Shaheen for books and NLLB 3.3B for STEM papers and Wiki Encyclopedia (Team et al., 18 Jan 2025).

Data quality control is a major part of the model’s construction. Records are standardized into JSON; cleaning removes HTML and JavaScript tags and normalizes whitespace; filtering proceeds through syntactic, semantic, and model-based stages; semantic filtering uses ASAD; model-based filtering uses KenLM perplexity filtering and an education classifier; and deduplication includes URL-level, exact-match, and approximate-match stages. For the approximate deduplication stage, the paper reports gram size $= 8$ , number of bands $b = 12$ , band length $r = 11$ , approximated Jaccard threshold $\approx 0.8$ , and signature length $= 132$ (Team et al., 18 Jan 2025).

3. Instruction tuning and alignment stack

The “Instruct” suffix is grounded in explicit post-training rather than deployment-time prompting alone. The platform paper states that the same post-training stages were applied to Fanar Star and Fanar Prime, and that both models were trained to support a system prompt as part of their chat template. The post-training pipeline consists of supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) for safety and value alignment (Team et al., 18 Jan 2025).

The SFT corpus combines three sources: curated public instruction/dialog datasets, synthetic data generation, and new capability data. The public-source curation process produced about 2.5 million instructions and dialogues across 11 capability categories in Arabic and English. Synthetic generation produced close to a million samples in both languages, mainly for cultural contextualization and value alignment, using stronger external instruct models such as Gemma-2-27B-it, Qwen2.5-72B-Instruct, c4ai-command-r-plus, Llama-3.1-70B-Instruct, and Llama-3.1-405B-Instruct (Team et al., 18 Jan 2025). Additional capability data covered closed-set QA behavior expansion, IslamQ&A, poetry, humor, diacritization, question generation, grammar correction, dialectal dialogues, and expert DPO data on nuanced and controversial Islamic issues (Team et al., 18 Jan 2025).

The SFT pipeline is two-stage. For Fanar Prime, SFT Stage 1 uses 3.6M samples, batch 640, and learning rate $5.0\times10^{-7}$ , while SFT Stage 2 uses 834K samples, batch 640, and learning rate $1.0\times10^{-7}$ (Team et al., 18 Jan 2025). Preference learning then adds 250K preference samples, balanced across Arabic and English with about 20% on-policy, and Table 8 reports batch 640 and learning rate $1.0\times10^{-7}$ for DPO (Team et al., 18 Jan 2025).

The Fanar 2.0 retrospective compresses the first-generation 9B alignment profile into a concise systems description: SFT + DPO, 8K tokens context, no native reasoning traces, no tool calling, translation-based traces, SFT-based cultural alignment, and knowledge probing as hallucination mitigation (Team et al., 17 Mar 2026). This makes the historical contrast with Fanar 2.0 explicit: Fanar-1-9B-Instruct is an aligned instruct model, but it predates later additions such as native Arabic reasoning traces, self-verification traces, longer context, and integrated tool calling.

4. Empirical profile across benchmark families

In the Fanar 2.0 comparison, Fanar-1-9B-Instruct functions as a broad predecessor baseline rather than a single-task model, and its reported scores span Arabic knowledge, Arabic language, dialectal Arabic, Arabic and Islamic culture, English world knowledge, and conversational instruction following (Team et al., 17 Mar 2026).

Evaluation family	Fanar-1-9B-Instruct	Fanar-27B
Arabic knowledge	MMMLU/Ar 58.30; ArabicMMLU 67.35; OALL-v2 68.64	67.40; 74.67; 69.40
Arabic language	Nahw-MCQ 40.00; AraLingBench 60.60	46.88; 68.67
Dialectal Arabic	Belebele 83.26; DialectalArabicMMLU 59.91	86.81; 67.40
Arabic and Islamic culture	ACVA 79.66; PalmX Islamic 82.33; PalmX Culture 67.10	82.70; 85.38; 72.70
English world knowledge	MMLU 71.32; PIQA 82.37; HellaSwag 83.01; ARC Challenge 65.19	78.89; 85.91; 85.32; 65.61
Conversational / instruction	MT-Bench 5.58; IFEval 74.70	6.12; 82.97

These results establish two points. First, the model is already strong on Arabic and Islamic culture, with 79.66 on ACVA and 82.33 on PalmX Islamic (Team et al., 17 Mar 2026). Second, the benchmark profile is bilingual rather than Arabic-only: the same model records 71.32 on MMLU, 82.37 on PIQA, and 83.01 on HellaSwag (Team et al., 17 Mar 2026).

Under the earlier Fanar Prime Instruct naming, the same 9B line is also reported as strong on Arabic automatic and generative evaluations, including MT-Bench(Ar) 8.93, User-Chats 9.66, Belebele(Ar) 82.48, OALL 63.81, ArabicMMLU(Egy) 59.22, and ArabicMMLU(Lev) 60.14. The platform paper summarizes the 9B instruct line as “a strong competitor in Arabic, producing top scores in many automated and generative tasks” (Team et al., 18 Jan 2025).

5. Downstream evaluation and task-specific adaptation

A direct downstream evaluation of Fanar-1-9B-Instruct appears in Arabic automatic essay scoring. On the QAES dataset, the model is evaluated under Level 1: standard prompting, Level 2: hybrid prompting, and Level 3: rubric-guided prompting with scored exemplars, across the traits Organization, Vocabulary, Style, Development, Mechanics, Structure, Relevance, and the summed Final Score (Mandhari et al., 20 Mar 2026). Its total QWK rises from 0.008 under Level 1 to 0.161 under Level 2 and 0.284 under Level 3, which the paper identifies as the best total score in the study (Mandhari et al., 20 Mar 2026).

The per-trait pattern is also informative. Under Level 3, Fanar-1-9B-Instruct reaches Development 0.204, Mechanics 0.280, Relevance 0.227, and Total 0.284, while its strongest Style result appears under Level 2 at 0.132 rather than under Level 3 (Mandhari et al., 20 Mar 2026). The study states that Fanar-1-9B-Instruct achieves the highest trait-level agreement in both zero- and few-shot prompting, reports QWK = 0.28 and CI upper bound = 0.41, and interprets the results as evidence that strong instruction tuning, robust Arabic coverage, and structured prompting matter more than parameter scale alone for Arabic AES (Mandhari et al., 20 Mar 2026).

A closely related but distinct 9B lineage appears in the QIAS 2025 Islamic inheritance reasoning system. That work fine-tunes the Fanar-1-9B causal LLM, not explicitly the instruct checkpoint, using LoRA over q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj, with rank $r = 32$ , LoRA alpha $\alpha = 64$ , dropout $b = 12$ 0, 4 epochs, batch size 2, gradient accumulation 32, learning rate $b = 12$ 1, weight decay 0.01, warmup ratio 0.1, cosine decay, and FP16 (AL-Smadi, 20 Aug 2025). Integrated into a RAG pipeline using all-MiniLM-L6-v2, FAISS, top- $b = 12$ 2 retrieval, maximum input length 10,000 tokens, maximum 15 new tokens, temperature 0.05, and greedy decoding, the adapted system reaches 85.8% overall accuracy and 97.6% on the Advanced subset (AL-Smadi, 20 Aug 2025). Although this result is not a direct evaluation of Fanar-1-9B-Instruct itself, it documents the broader adaptability of the first-generation 9B Fanar line in a highly structured Arabic-Islamic reasoning domain.

6. Ambiguities, misconceptions, and scope of claims

Two recurrent ambiguities structure the literature on this model. First, the foundational Fanar paper does not use the exact string “Fanar-1-9B-Instruct”; it uses Fanar Prime and Fanar Prime Instruct (Team et al., 18 Jan 2025). Second, not every later reference to Fanar-1-9B-Instruct is reliable evidence about the 9B instruct model. In the CultranAI PalmX submission paper, the abstract says that Fanar-1-9B-Instruct achieved the highest performance, but the detailed notes report that the paper’s tables and appendix repeatedly refer to Fanar-7B / Fanar-7B-Instruct, and that the most accurate reading is that the model actually benchmarked, fine-tuned, and submitted is Fanar-7B-Instruct rather than the 9B model (Bhatti et al., 24 Aug 2025).

System-level claims likewise require disentanglement from checkpoint-level claims. Fanar-Sadiq is a bilingual multi-agent Islamic assistant and a core component of the Fanar AI platform; it states that the system uses “Fanar” as the LLM agent and compares against Fanar-2-27B, but it does not explicitly mention Fanar-1-9B-Instruct anywhere (Abbas et al., 9 Mar 2026). Its reported performance therefore characterizes an agentic system with routing, retrieval, deterministic calculators, citation normalization, verification traces, exact verse lookup, and tool use, not a standalone 9B instruct checkpoint (Abbas et al., 9 Mar 2026).

The first-generation 9B line also has substantive technical limits in the primary Fanar papers. Some core architectural items remain unspecified in the original platform description, including the exact number of layers, attention-head count, intermediate size, and exact context-length entry for Fanar Prime (Team et al., 18 Jan 2025). The same paper reports a dislike rate around 13–15%, attributes most dislikes to factuality-related errors, and introduces Recency RAG, Islamic RAG, Biography RAG, and Attribution RAG precisely because recency, verification, and some forms of grounded answering are not native model properties (Team et al., 18 Jan 2025). It also states that the platform has multimodal capabilities, but that these capabilities are not intrinsic to the autoregressive model (Team et al., 18 Jan 2025). This suggests that Fanar-1-9B-Instruct is best understood as a strong first-generation Arabic-centric instruct model whose deployed behavior was often coupled to a larger orchestration, retrieval, and safety stack.