olivilo/ zora:v1.11

36 Downloads Updated 1 week ago

Open, honest 8B LLM for 12 Balkan & SE-European languages (sr, hr, bs, mk, sl, sq, cnr, bg, el, tr, ro, hu). Thinks in-language, admits what it doesn't know. #1 on BalkanBench — beats models 3–4× its size. By Sovasoft.

ollama run olivilo/zora:v1.11

curl http://localhost:11434/api/chat \
  -d '{
    "model": "olivilo/zora:v1.11",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='olivilo/zora:v1.11',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'olivilo/zora:v1.11',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 1 month ago

1 month ago

b68e021debbe · 5.9GB ·

model

archqwen3

·

parameters8.19B

·

quantizationQ5_K_M

5.9GB

params

{ "stop": [ "<|im_end|>", "<|im_start|>" ] }

59B

Readme

Zora v1.11 — Previous Stable

зора = “dawn”. One to unite them all. — by Sovasoft (ai.in.rs)

An open 8B LLM for 12 Balkan & SE-European languages. Built on Qwen3-8B with in-language thinking, multi-perspective reasoning, and honest uncertainty.

BalkanBench: ⁸⁴⁄₁₅₆ — #1 among 8B models at release.

Note: v1.11 has been superseded by v1.12 which adds IDK Calibration, Tool-Cascade, and +4 points on BalkanBench (85). Recommended for new projects.

Quick Start

ollama run olivilo/zora-v1.11

Quantizations

Tag	Type	Size	Notes
`latest`	Q5_K_M	5.0 GB	Recommended
`q6`	Q6_K	6.7 GB	Near-lossless
`q8`	Q8_0	8.7 GB	Highest fidelity

Avoid Q4 and below — caused hallucination in our tests.

How It Thinks

In-Language — thinks in the language you write in
Multi-Perspective — shows different sides on contested topics
Honest — admits uncertainty instead of guessing
Multi-lingual — 12 languages: sr, hr, bs, mk, sl, sq, cnr, bg, el, tr, ro, hu

Benchmark

Model	Size	Score
Zora v1.11	8B	84
Gemma-4-31B	31B	77
Mistral-24B	24B	73
Qwen3.6-30B	30B	73
Salamandra	7B	66
EuroLLM	9B	65
Aya	8B	61
BgGPT	7B	56

Full charts (13 languages): huggingface.co/sovasoft/zora-v1.11

Limits

Multi-step logic/math weak (8B limit) — improved in v1.12
Factual depth limited — use RAG for exact facts
English decent but not state-of-the-art

Links

License: Apache-2.0 | Base: Qwen3-8B | Context: 8K | Languages: sr, hr, bs, mk, sl, sq, cnr, bg, el, tr, ro, hu

<img src="https://huggingface.co/sovasoft/zora-v1.11/resolve/main/images/Zora_Stich_1700_SW_Print.png" width="420"/>

# Zora v1.11 — Previous Stable

*зора = "dawn". One to unite them all.* — by **Sovasoft** ([ai.in.rs](https://ai.in.rs))

An open 8B LLM for **12 Balkan & SE-European languages**. Built on Qwen3-8B with in-language thinking, multi-perspective reasoning, and honest uncertainty.

**BalkanBench: 84/156** — #1 among 8B models at release.

---

> **Note:** v1.11 has been superseded by [v1.12](https://ollama.com/olivilo/zora-v1.12) which adds IDK Calibration, Tool-Cascade, and +4 points on BalkanBench (85). Recommended for new projects.

---

## Quick Start

```bash
ollama run olivilo/zora-v1.11
```

### Quantizations

| Tag | Type | Size | Notes |
|---|---|---|---|
| `latest` | Q5_K_M | 5.0 GB | **Recommended** |
| `q6` | Q6_K | 6.7 GB | Near-lossless |
| `q8` | Q8_0 | 8.7 GB | Highest fidelity |

**Avoid Q4 and below** — caused hallucination in our tests.

---

## How It Thinks

- **In-Language** — thinks in the language you write in
- **Multi-Perspective** — shows different sides on contested topics
- **Honest** — admits uncertainty instead of guessing
- **Multi-lingual** — 12 languages: sr, hr, bs, mk, sl, sq, cnr, bg, el, tr, ro, hu

---

## Benchmark

| Model | Size | Score |
|---|---|---|
| **Zora v1.11** | **8B** | **84** |
| Gemma-4-31B | 31B | 77 |
| Mistral-24B | 24B | 73 |
| Qwen3.6-30B | 30B | 73 |
| Salamandra | 7B | 66 |
| EuroLLM | 9B | 65 |
| Aya | 8B | 61 |
| BgGPT | 7B | 56 |

Full charts (13 languages): [huggingface.co/sovasoft/zora-v1.11](https://huggingface.co/sovasoft/zora-v1.11)

---

## Limits

- Multi-step logic/math weak (8B limit) — improved in v1.12
- Factual depth limited — use RAG for exact facts
- English decent but not state-of-the-art

---

## Links

- [Model Card + Charts](https://huggingface.co/sovasoft/zora-v1.11)
- [BalkanBench](https://github.com/olivilo/balkanbench)
- [All Zora models](https://ollama.com/olivilo/zora)

**License:** Apache-2.0 | **Base:** Qwen3-8B | **Context:** 8K | **Languages:** sr, hr, bs, mk, sl, sq, cnr, bg, el, tr, ro, hu

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)