Zora v1.12 (MLX) — Open, honest LLM for the Balkans

zora = "dawn". The one that unites them all. — by Sovasoft (ai.in.rs)

This is the MLX version of Zora v1.12, optimized for Apple Silicon Macs (M1/M2/M3/M4).

What is Zora

Zora is an open 8B language model (built on Qwen3-8B) for 12 Balkan languages: Serbian, Croatian, Bosnian, Macedonian, Slovenian, Albanian, Montenegrin, Bulgarian, Greek, Turkish, Romanian, Hungarian.

What is MLX

MLX is Apple's machine learning framework for Apple Silicon. This version runs natively on your Mac's GPU with:

Faster inference than PyTorch/Transformers
Lower memory usage through MLX optimization
Native Apple Silicon support (M1, M2, M3, M4)

Install

pip install mlx mlx-lm

Usage

Python

from mlx_lm import load, generate

model, tokenizer = load("sovasoft/zora-v1.12-mlx")

messages = [{"role": "user", "content": "Ko si ti?"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=prompt, max_tokens=512)
print(response)

Command Line

mlx_lm.generate --model sovasoft/zora-v1.12-mlx --prompt "Ko si ti?"

Server

mlx_lm.server --model sovasoft/zora-v1.12-mlx --host 0.0.0.0 --port 8080

Benchmark

BalkanBench: 85/156 (v1.12, 16-bit). See the main model card for full details.

Quantize (optional)

For lower memory usage:

mlx_lm.convert --hf-path sovasoft/zora-v1.12 --mlx-path zora-v1.12-mlx-q4 --q-bits 4

License

Apache-2.0

Acknowledgements

Thanks to the Qwen team at Alibaba for Qwen3 (Apache-2.0), and to the MLX team at Apple for the MLX framework.

зора — the dawn belongs to everyone.

Sovasoft · ai.in.rs · one to unite them all

Downloads last month: 32

Safetensors

Model size

8B params

Tensor type

BF16

MLX

Hardware compatibility

Quantized

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for sovasoft/zora-v1.12-mlx

Base model

Qwen/Qwen3-8B-Base

Finetuned

Qwen/Qwen3-8B

Finetuned

(1973)

this model