Persian AI in 2026: The Best Local Models for Farsi Speakers

What Changed in 2026

Two shifts define Persian AI this year. First, evaluation caught up with the language. Instead of translating English MMLU and hoping, researchers released Persian-native benchmarks: DivanBench for cultural reasoning (EACL 2026 SilkRoadNLP), Parse for open-domain reasoning QA (SIGIR 2026), PMWP for math word problems, and APARSIN for sentiment and translation across 14 Iranic languages and dialects.

Second, the open-weight frontier moved. The 2026 generation — Qwen3.5, Gemma 3, gpt-oss-20b, Tiny-Aya-Global — is materially better at Persian than the models that dominated 2024, and most of it runs offline on ordinary hardware. The practical question is no longer "can a local model handle Farsi?" but "which current model, for which task?"

The 2026 Model Lineup

These are the current-generation models with Persian results reported in 2026 research. Scores are quoted from the papers cited at the bottom of this page.

Persian Reasoning in 2026

Parse (SIGIR 2026) is the first open-domain reasoning QA benchmark built for Persian — 10,800 questions across Boolean, multiple-choice and factoid formats, including multi-hop and unanswerable cases. Its most useful practical findings for anyone building on Persian today are about method, not a single leaderboard score: Persian-language prompts outperform English ones, and structured prompting helps — chain-of-thought for Boolean and multiple-choice items, few-shot for factoid questions.

The takeaway for the 2026 generation: reasoning-heavy Persian work still rewards scale and prompting discipline, and it is worth re-running Parse against whichever current model you deploy rather than assuming a ranking carries over.

Persian-Specific Models in 2026

The dedicated Persian track is alive and still worth knowing. Current Persian-optimized open models include Dorna / Dorna2 (Part AI), Maral, PersianMind, and Hormoz-8B. The Awesome-Persian-LLM repository curates the ecosystem — models, datasets, papers and tooling in one place.

Two honest caveats, both supported by 2026 research. First, Persian-specific pretraining does not automatically improve reasoning: DivanBench showed that adapting a base model to Persian raised acceptance of cultural statements while cutting its ability to reject violations. Second, dedicated Persian models tend to win on fluency and terminology, while general multilingual models win on reasoning breadth. Use specialists for language and format, generalists for thinking.

How to Choose in 2026

Running It Locally

The 2026 lineup is friendlier to local hardware than the size numbers suggest. Roughly 0.5GB of memory per billion parameters for quantized weights means a 9B model fits comfortably in 8GB, and a 12B model in 16GB. So Qwen3.5-9B and Gemma 3 12B both run fully offline on a modern laptop; gpt-oss-20b wants ~24GB class hardware. For sizing, see How Much RAM Do You Need to Run Local AI? and Top 10 GGUF Models Ranked by RAM.

Three Techniques That Actually Improve Persian AI

1. Fine-tune on clean Persian data. PMWP 2026 showed the point vividly: after LoRA fine-tuning, small open models jumped past 91% exact equation match on Persian math. A small, curated dataset beats a bigger generic model — see Fine-Tuning a Local LLM: LoRA for Beginners.
2. Lock the glossary. For legal, medical or government terminology, a glossary-constrained pipeline keeps translation consistent — still the highest-leverage technique in regional languages. See Building a Multilingual Translation Pipeline with Local LLMs.
3. Evaluate in Persian, with scenario tests. Use DivanBench, Parse, PMWP, APARSIN, MELAC, ParsiNLU, Persian MMLU and FaMTEB. Check both recall and rejection — the acquiescence trap means a model that looks accurate on facts may still fail on judgment.

Why Local Still Matters

Cloud models keep a narrow edge on some Persian tasks, but the 2026 open generation has closed most of the gap for everyday work — and it comes without the strings: no subscription, no data leaving the device, no availability problems for speakers in restricted environments. Local Persian AI is free, private, unlimited and offline. For institutional work — legal, medical, government — it is often the only acceptable option, and it puts the capability in the hands of the speakers themselves.

Frequently Asked Questions (FAQ)

Which model is best for Persian in 2026?

It depends on the task. On DivanBench (2026), Gemma 3 12B led factual Persian knowledge (87%) and cultural reasoning (66%) among 7–12B models. On the MEA multilingual benchmark (2026, 24 languages including Persian), Qwen3.5-9B was the strongest model overall. For Persian math word problems (PMWP, 2026), Gemini 2.5 Flash scored highest (72.02%).

Are there 2026 models built specifically for Persian?

Persian-specialized open models continue to be developed and adapted — Dorna2 is a current Persian-optimized model evaluated in 2026 research, alongside Maral, PersianMind and Hormoz. Most 2026 progress, however, comes from general multilingual models like Qwen3.5 and Gemma 3, whose Persian capability improves with each generation.

Can I run a 2026 Persian model locally?

Yes. Qwen3.5 ships in 2B and 9B sizes and Gemma 3 in a 12B size — both fit quantized on consumer hardware (roughly 8–16GB RAM for the 7–12B class). gpt-oss-20b and Tiny-Aya-Global are also open weights. All run fully offline and private.

How should I evaluate Persian AI in 2026?

Use Persian-native benchmarks rather than English ones. The 2026 set includes DivanBench (cultural reasoning), Parse (open-domain reasoning QA), PMWP (math word problems), APARSIN (Iranic sentiment and translation), plus MELAC, ParsiNLU, Persian MMLU and FaMTEB.

Is Persian AI better in the cloud?

Cloud models such as Gemini 2.5 Flash still lead some Persian tasks, but open 2026 models like Qwen3.5 and Gemma 3 are close enough for most work — and they are private, free, unlimited and offline. For sensitive or institutional Persian work, local deployment is often the only acceptable option.