Persian quality on your work
Public tables are only a starting point. We test several models on real samples of your work, such as questions, documents and letters, and go through the results with you.
The local AI models we run on your servers
Any model whose weights are published runs on your own hardware, from China's DeepSeek, Qwen, Kimi and GLM to Gemma and Mistral. We choose the model after testing several on your own Persian work, with a licence that covers your commercial use.
What a local model is
A local (open-weight) model is one whose maker has published the weight files so anyone can run it on their own hardware, unlike ChatGPT or Claude, which are available only through their makers' services. Your data never leaves your server and an international internet cut does not stop the work. Before choosing we check two things: the model's Persian quality on your own work, and the terms of its licence for commercial use.
How we choose a model
Size against hardware
A smaller model that answers quickly and correctly beats a large one that barely fits and runs slowly.
Commercial licence
We read every model's licence. Some are free (MIT and Apache 2.0), some set conditions for large-scale serving or very large companies, and some are non-commercial only.
Serving software
We check that the model runs on vLLM, llama.cpp or MLX; some new models are not supported everywhere yet.
The best open models on a Persian test
The Open Persian LLM Leaderboard is the only current Persian benchmark for language models: 21 tasks, from Persian literature and proverbs to law and Iranian social norms (data up to 19 September 2026). The simple mean of the 21 tasks is our calculation, not the board's official rank.
| Model | Maker | Persian test mean | Size | Suitable hardware |
|---|---|---|---|---|
| Gemma 4 31B | 78.6, above GPT-4o at 76.3 | 31B parameters | One RTX 5090 at 4-bit, or one Mac Studio | |
| Qwen3.8-27B | Alibaba (Qwen) | 76.1, the best Persian-literature score | 28B parameters | One RTX 5090 at 4-bit, or one Mac Studio |
| Gemma 4 26B-A4B | 70.7 | 26B, 4B active | One 24 to 32 GB card, DGX Spark or Ryzen AI Max | |
| Llama 4 Scout | Meta | 69.0 | 109B, 17B active | One RTX PRO 6000 or a Mac Studio M5 Max |
| Gemma 4 E4B | 59.0 | About 4.5B effective | A Mac mini or one 8 GB card |
Cohere's Command A+ lists Persian among its 48 training languages, and its predecessor Command A scored 70.0 on the same board. For every other model, including all the large Chinese models, we found no published Persian score; we test them on your own work.
Large models: 200 billion parameters and up
Size, licence and date from each model's official Hugging Face card, reviewed on 22 October 2026. "Active" is the parameters used for each word; in MoE models it sets the speed, while the total sets the memory needed.
| Model | Maker | Total / active | Licence | Memory at 4 / 8-bit | Suitable hardware |
|---|---|---|---|---|---|
| DeepSeek-V4.1-Flash | DeepSeek, China | 552B plus 196B Engram / 8B to 16B | MIT | About 460 / 920 GB (official files about 510) | An eight-GPU H200 or B200 server |
| DeepSeek-V4-Pro | DeepSeek, China | 1.6T / 49B | MIT | About 990 / 1,980 GB | An eight-GPU B200 server or larger, or a Mac Studio cluster |
| DeepSeek-V4-Flash | DeepSeek, China | 284B / 13B | MIT | About 170 / 340 GB | One Mac Studio M5 Ultra or a four-GPU server |
| Kimi K3 | Moonshot AI, China | 2.8T / 104B | Kimi K3 License, with conditions for large-scale serving | About 1,670 / 3,340 GB | An eight-GPU B300 or MI355X server at 4-bit |
| Kimi K2.7 Code | Moonshot AI, China | 1T / 32B | Modified MIT | About 620 / 1,230 GB | An eight-GPU H200 server, or two 512 GB Mac Studios |
| Qwen3.8-2.4T-A95B | Alibaba, China | 2.4T / 95B | Qwen3.8-Max License, with conditions for large-scale serving | About 1,470 / 2,940 GB | An eight-GPU B300 or MI355X server |
| Qwen3.5-397B-A17B | Alibaba, China | 397B / 17B | Apache 2.0 | About 240 / 480 GB | A 512 GB Mac Studio or a four-GPU H200 server |
| GLM-5.2 and GLM-5.3 | Z.ai, China | 744B / 40B | GLM-5.2: MIT; GLM-5.3: conditions only for very large companies | About 450 / 900 GB | An eight-GPU H200 server, or a 512 GB Mac Studio at 4-bit |
| MiMo-V2.6 (Pro and Flash) | Xiaomi, China | Pro: 1T / 42B; Flash: 309B / 15B | MIT | Pro about 610 / 1,230; Flash about 190 / 370 GB | Pro: an eight-GPU server; Flash: one Mac Studio M5 Ultra |
| MiniMax-M3 | MiniMax, China | About 428B / 23B | MiniMax Community License; commercial use with notice and "Built with MiniMax M3" | About 256 / 512 GB | A 512 GB Mac Studio or a four-GPU H200 server |
| Tencent Hy3 | Tencent, China | 295B / 21B | Apache 2.0 | About 180 / 360 GB | One Mac Studio M5 Ultra or a four-GPU server |
| Ling-3.1-flash | Ant Group, China | 560B / 25B | MIT | About 340 / 690 GB | An eight-GPU H200 server |
| Muse Spark Open | Meta, US | 816B / 42B | Muse licence; "Built with" notice and conditions for large-scale serving | About 490 / 990 GB | Four Mac Studio M5 Ultras (Apple's test) or an eight-GPU server |
| Nemotron 3 Ultra | NVIDIA, US | 550B / 55B | OpenMDW-1.1 (permissive) | About 340 / 670 GB | An eight-GPU H200 or B200 server |
| Mistral Large 3 | Mistral, Europe | 675B / 41B | Apache 2.0 | About 405 / 810 GB | An eight-GPU H200 server |
| Command A+ | Cohere, Canada | 218B / 25B | Apache 2.0 | About 130 / 260 GB | Two H100 cards or one Mac Studio M5 Ultra |
The memory column is our estimate of the weights alone (total parameters x bits / 8 x 1.2); long context and concurrent users need more. Many official files are already 8-bit or 4-bit.
Mid-size models: 20 to 130 billion parameters
| Model | Maker | Total / active | Licence | Memory at 4 / 8-bit | Suitable hardware |
|---|---|---|---|---|---|
| Qwen3.8-27B | Alibaba, China | 28B | Apache 2.0 | About 17 / 33 GB | One RTX 5090 or a Mac mini M5 Pro |
| Qwen3.6-35B-A3B | Alibaba, China | 36B / 3B | Apache 2.0 | About 22 / 43 GB | One RTX 5090 or a Mac mini M5 Pro |
| Qwen3.5-122B-A10B | Alibaba, China | 125B / 10B | Apache 2.0 | About 75 / 150 GB | One RTX PRO 6000, DGX Spark or a Mac Studio M5 Max |
| Gemma 4 31B and 26B-A4B | Google, US | 31B; 26B / 4B | Apache 2.0 | About 19 / 38 and 15 / 31 GB | One RTX 5090 or a Mac Studio |
| gpt-oss-120b and gpt-oss-20b | OpenAI, US | 117B / 5B; 21B / 4B | Apache 2.0 | About 70 / 140 and 13 / 25 GB | 120b: RTX PRO 6000, DGX Spark or Ryzen AI Max; 20b: one 16 GB card |
| Mistral Small 4 | Mistral, Europe | 119B / 6.5B | Apache 2.0 | About 71 / 143 GB | One RTX PRO 6000 or a Mac Studio M5 Max |
| Mistral Medium 3.5 | Mistral, Europe | 128B | Modified MIT; no rights for companies above US$20M monthly revenue | About 77 / 154 GB | Two RTX PRO 6000 or a Mac Studio M5 Ultra |
| Muse Glimmer 30B | Meta, US | 30B | Apache 2.0 | About 18 / 36 GB | One RTX 5090 or a Mac Studio |
| Nemotron 3.5 Super VL | NVIDIA, US | 120B / 12B | OpenMDW-1.1 | About 74 / 149 GB | One RTX PRO 6000 or DGX Spark |
| GLM-4.7-Flash | Z.ai, China | 30B / 3B | MIT | About 19 / 37 GB | One RTX 5090 or a Mac mini M5 Pro |
| Seed-OSS-36B | ByteDance, China | 36B | Apache 2.0 | About 22 / 43 GB | One RTX 5090 or RTX PRO 5000 |
| ERNIE-4.5-21B-A3B | Baidu, China | 22B / 3B | Apache 2.0 | About 13 / 26 GB | One 24 GB card |
| Kolibri-1 | Aleph Alpha, Germany | 78B / 3.5B | Apache 2.0 | About 47 / 94 GB | One RTX PRO 6000 or a Mac Studio M5 Max |
| Granite 4.2 30B | IBM, US | 29B | Apache 2.0 | About 18 / 35 GB | One RTX 5090 |
Small models: up to 15 billion parameters
For laptops, a Mac mini, small computers, and simple high-volume jobs such as classification, short summaries and data extraction.
| Model | Maker | Size | Licence | Memory at 4 / 8-bit | Note |
|---|---|---|---|---|---|
| Gemma 4 12B | Google, US | 12B; text, image and audio | Apache 2.0 | About 7 / 14 GB | Trained on more than 140 languages |
| Gemma 4 E4B and E2B | Google, US | About 4.5B and 2.3B effective | Apache 2.0 | About 5 / 10 and 3 / 6 GB | Persian test 59.0 and 50.5 |
| Qwen3.5-9B, 4B and 2B | Alibaba, China | 9.7B, 4.7B and 2.3B | Apache 2.0 | From about 1.4 to 12 GB | Text and image |
| MiMo-V2.6-Distill-Qwen-9B | Xiaomi, China | 9.4B | MIT | About 6 / 11 GB | A distilled version of Xiaomi's MiMo |
| Ministral 3 (14B, 8B and 3B) | Mistral, Europe | 14B, 9B and 4B | Apache 2.0 | From about 2 to 17 GB | Text and image |
| Phi-4 and Phi-4-mini | Microsoft, US | 15B and 3.8B | MIT | About 9 / 18 and 2 / 5 GB | Reasoning and maths |
| Granite 4.2 (8B and 3B) | IBM, US | 8.8B and 3.7B | Apache 2.0 | About 5 / 11 and 2 / 4 GB | Light enterprise tasks |
| MiniCPM5-2B | OpenBMB, China | 2.5B | Apache 2.0 | About 1.5 / 3 GB | For small devices |
Coding models
| Model | Maker | Total / active | Licence | Memory at 4 / 8-bit | Suitable hardware |
|---|---|---|---|---|---|
| Qwen3-Coder-Next | Alibaba, China | 80B / 3B | Apache 2.0 | About 48 / 96 GB | One RTX PRO 6000 or a Mac Studio M5 Max |
| Qwen3-Coder-30B-A3B and 480B-A35B | Alibaba, China | 30B / 3B; 480B / 35B | Apache 2.0 | About 18 / 37 and 288 / 576 GB | 30B: one RTX 5090; 480B: a 512 GB Mac Studio or an eight-GPU server |
| Kimi K2.7 Code | Moonshot AI, China | 1T / 32B | Modified MIT | About 620 / 1,230 GB | An eight-GPU H200 server or two 512 GB Mac Studios |
| Devstral Small 2 | Mistral, Europe | 24B | Apache 2.0 | About 14 / 29 GB | One RTX 5090 |
| Devstral 2 | Mistral, Europe | 125B | Modified MIT; no rights for companies above US$20M monthly revenue | About 75 / 150 GB | Two RTX PRO 6000 or a Mac Studio M5 Ultra |
Qwen3.8-27B, GLM-5.3 and DeepSeek-V4 are strong at coding too; we compare them for self-hosting in a separate article.
Speech, images, documents and search
| Job | Model | Licence | Persian status |
|---|---|---|---|
| Speech to text, batch | Whisper large-v3 and v3-turbo, with Persian fine-tunes | Apache 2.0 and MIT; Persian fine-tunes CC-BY-4.0 or MIT | Stock Whisper has about 36% word error on the FLEURS Persian test; Persian fine-tunes from October 2026 report much lower error, by their makers' own figures |
| Speech to text, live | nemotron-asr-streaming-farsi, based on an NVIDIA model | OpenMDW-1.1 | 8.8% word error on the full FLEURS Persian test, as reported by its maker; Persian only |
| Speech to text, multilingual | Meta Omnilingual ASR | Apache 2.0 | 3.4% character error for Persian in Meta's official table |
| Text to speech | Piper (fa_IR voices) and MOSS-TTS v1.5 | MIT and Apache 2.0 | The only Persian options with a free commercial licence; the best-measured Persian voices (OmniVoice and Higgs) are non-commercial only |
| Reading images and documents (OCR) | Qwen3-VL, DeepSeek-OCR-2, PaddleOCR-VL and GLM-OCR | Apache 2.0 and MIT | None publishes Persian OCR accuracy; we test on your own scans |
| Search over documents (embeddings) | bge-m3, multilingual-e5-large and Nemotron-3.5-Embed | MIT and OpenMDW-1.1 | bge-m3 scores 59.1 on the Persian FaMTEB benchmark; Nemotron-3.5-Embed lists Persian among its evaluated languages |
| Image generation | FLUX.2 klein 4B, Qwen-Image-2512 and ERNIE-Image | Apache 2.0 | None guarantees Persian text inside images |
Speech error figures come from different sources and test subsets and are not comparable with each other; we test every speech model on your own audio.
From the first call to a working system
Hardware is chosen after a model has been measured on your own work, not from a table. A quote is issued only once it is clear which model has to serve how many people, and how fast.
- 01
First conversation
We go through the task, the number of users, the data, where the hardware goes and what you already own.
- 02
Measured on your work
We test several open-weight models on samples of your real work and size memory and speed from those results.
- 03
Written quote
You get the exact configuration, the proposed model, the delivery time and the support terms in writing. Prices appear only in the quote.
- 04
Installation
We install the hardware, bring up the model and connect the assistant to your systems, then test it once with international internet cut off.
- 05
Handover and support
You get documentation and training for your team; afterwards we maintain the system and update the models under contract.
How to get in touch
Fill in the form or message us on WhatsApp. The clearer these are, the more exact the quote:
- What the model has to do (answer from documents, code, summarise, process images or speech)
- How many people use it at the same time, and what response time is acceptable
- The model or hardware family you have in mind, if you already know
- Where it goes: your server room, a data centre, or hardware you already own
- How sensitive the data is, and whether it must run with no internet at all
We reply within one business day.
Zharf
Try the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about local models
Which model is best for Persian?
On the public Open Persian LLM Leaderboard, Gemma 4 31B (mean 78.6) and Qwen3.8-27B (76.1) are the best open models, and both run on one 32 GB card or one Mac Studio. The large Chinese models such as DeepSeek and Kimi have no published Persian score. We make the final choice by testing on your own work.
What hardware does running DeepSeek on our own server need?
It depends on the version. DeepSeek-V4-Flash, at 284 billion parameters, has about 170 GB of files and runs on one Mac Studio M5 Ultra or a four-GPU server. DeepSeek-V4.1-Flash is about 510 GB and needs an eight-GPU H200 or B200 server. DeepSeek-V4-Pro, at 1.6 trillion parameters, needs larger servers or a cluster.
Is a local model as good as ChatGPT or Claude?
Not always. The largest open models come close to closed models on some benchmarks, but a model that fits on affordable hardware is usually smaller. For defined jobs such as answering from documents, summarising and extracting data, a local model is often enough; we measure it on your real work and say plainly what we find.
Are open-model licences safe for commercial use?
Most of the models we recommend have free MIT or Apache 2.0 licences. Some, such as Kimi K3, Muse, MiniMax-M3 and Mistral Medium 3.5, set conditions for large-scale serving or very large companies, and a few image and speech models are non-commercial only. We read the licence before recommending a model and write it into the quote.
How are models updated?
We hand over the model installed and keep its version fixed, so the system's behaviour never changes without notice. When a better model is released, we test it on the same set of your tasks and replace the old one through an internal route only if it is better.
How much does it cost?
We do not publish prices, because they depend on the configuration, quantity, the day's market and how it is delivered. After the first conversation and a test on your work you receive a written quote that also names the exact hardware model, the delivery time and the support terms.
Does it work without international internet?
Yes, if the model, the serving software, the assistant, user sign-in and monitoring all sit inside your network. Anything that depends on a foreign service stops during a cut, so before handover we test the system with international internet deliberately cut off.
Start with the job the model has to do
Send a few samples of your real work. We test several models on them and tell you which model, on what hardware, under which licence; and if no local model is good enough, we say so plainly.