ibnsina-1.5b

ibnsina-1.5b

A small Persian language model for writing, summarizing, translating and chatting offline, trained from scratch mostly on Persian text. Runs through ollama, LM Studio or llama.cpp on a laptop.

Made by ibnsina-llm on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: ibnsina-1.5b is a small language model, about 1.5B parameters, for writing, summarizing, translating and chatting in Persian. Its authors say it was trained from scratch on mostly Persian text instead of adapting an English base, then tuned for conversation. It ships as GGUF files that run offline on your own device.

Who it suits: Persian speakers, developers and researchers who want local Persian text generation on modest hardware.

Industry | Who uses it | How they use it | Example
Content | Persian writers | Draft, summarize and translate Persian text offline | A short summary of a long Persian article
Software | App developers | Embed a small local Persian chat model | An offline assistant on a laptop
Research | NLP researchers | Study a Persian-first model and its public recipe | Compare it with the 3B sibling on the card's test sets

Anyone who needs answers about people, politics or news, or help with math and code, can skip it.

Where to use it: It runs locally through three tools, and the smallest file suits laptops and phones.

• ollama on Mac, Windows or Linux
• LM Studio
• llama.cpp, with the GGUF files
• Laptops and phones, using the smaller Q4_K_M file

Use case | Who | What they get back
Persian drafting | Writers | Fluent Persian text that you should check yourself
Summarizing and translating | Students and office staff | Shorter or translated Persian text
Offline chat | Developers | A conversation with no cloud service involved

How to set it up: Install one of the three runtimes, then pull or download the model.

• ollama, LM Studio or llama.cpp installed on your computer
• About 0.9 GB of disk space for the laptop and phone file
• Python with pip, only if you download the files with hf

1. Install ollama for your system.
2. Run the ollama command that pulls the model from Hugging Face.
3. Or search for the model name in LM Studio.
4. Or, for llama.cpp, download a GGUF from the repository and point llama-cli at it.

ollama run hf.co/ibnsina-llm/ibnsina-1.5b

pip install -U huggingface_hub
hf download ibnsina-llm/ibnsina-1.5b

The repository holds four files, so choose the one that fits your device.

File | Use
ibnsina-1.5b-Q4_K_M.gguf | Laptops and phones, about 0.9 GB
ibnsina-1.5b-Q8_0.gguf | Close to lossless, about 1.6 GB
ibnsina-1.5b-f16.gguf | Full precision, about 3 GB
Modelfile | The setup file for ollama

How to use it: Chat in ollama or LM Studio, or run llama-cli with a GGUF file. The chat template is stored inside the GGUF, and any system message is merged into the opening user message.

llama-cli -m ibnsina-1.5b-Q4_K_M.gguf
ollama create ibnsina-1.5b -f Modelfile && ollama run ibnsina-1.5b

Tool | What it does
ollama run | Pulls the model from Hugging Face and opens a chat
ollama create with the Modelfile | Builds a local model from the repository's Modelfile
llama-cli -m | Runs a GGUF file with llama.cpp

Example: You paste a Persian paragraph into the chat and ask for a two-sentence summary. You get fluent Persian text back. The card warns that such text can be wrong, so check any detail that matters before you use it.

For newcomers and experts: This is a small chatbot that writes in Persian and runs on your own computer, with no internet needed after the download. The first small thing to try is installing ollama, running the model command and asking for a short Persian paragraph.

The architecture is Llama-compatible, with 28 layers, width 2048, 16 heads with 4 KV heads, SwiGLU 6144, RMSNorm and RoPE theta 500k. Context is 2048 tokens, and the tokenizer is a 32,768-token byte-level BPE trained on mostly Persian text. Fine-tuning used 51.5k judged conversations across 17 categories.

Pretrain on mostly Persian text
Fine-tune on judged Persian conversations
Export the model as GGUF files
Chat through ollama, LM Studio or llama.cpp

Good to know: The card is frank about the model's size, and its scores sit near chance on multiple-choice tests.

Title: ParsiNLU-MC accuracy
ibnsina-1.5b: 31.8%
IbnSina-3B: 32.4%
Random guess: 25%

Title: Tokens per Persian word, lower is better
IbnSina: 1.29
Gemma 3: 1.64
Qwen3.5: 1.73

• Reliability: it can write fluent but wrong sentences, so verify anything that matters.
• Baseline: multiple-choice scores near 25 to 40 percent are close to chance, so treat them as a starting point.
• License: Apache-2.0 for the weights, the code and the build recipe. The training data includes only sources whose licenses allow that release.
• Cost to run: none named.
• Data it reaches: it generates text offline on your device. The files come from Hugging Face when you download them.
• Not for: looking up facts on public figures or current events, giving advice, answering knowledge quizzes, solving math or writing code. A bigger model suits those jobs.

On Hugging Face, 13 November 2026

• Task: text generation
• Library: llama.cpp
• Parameters: 1.5B
• Weights: about 742 MB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 803
• Likes: 3
• License: apache-2.0
• Languages: fa, en
• Tags: llama.cpp, gguf, persian, farsi, ibnsina, llama, chat, text-generation, conversational

Links

• Model: https://huggingface.co/ibnsina-llm/ibnsina-1.5b
• Files: https://huggingface.co/ibnsina-llm/ibnsina-1.5b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download ibnsina-llm/ibnsina-1.5b

Then follow the usage in its model card. Nothing in this kit runs by itself.