🦬 Zubr Mini 1.9 8B

Repository: Zubr-AI/zubr-mini-1.9-8b

---

📖 What This Model Is

zubr-mini-1.9-8b is a multimodal (Vision) language model with 8 billion parameters, fine-tuned from Mistral 3 8B (Ministral-3-8B-Instruct-2512 by Mistral AI). The author is Zubr-AI (alekringtonnn-ai). It is the flagship version of the Zubr lineup: it shares the same principles as the 3B version, but with noticeably better reasoning and image processing quality.

The model is distributed in GGUF format — a universal format for quantized weights designed for local inference. It runs on regular PCs (CPU), laptops, and servers, and even faster with a GPU (NVIDIA/AMD).

🧠 Key Feature: Fast Thinking

The model was trained using a special principle: each training example contained context, a question, and the beginning of small reasoning steps (short chains of thought).

Because of this training format, the model knows how to think fast — the reasoning process happens directly at the answer generation stage, rather than as a separate lengthy "thinking" phase. Unlike classic reasoning models that first unfold long chains of thought and only then answer, zubr-mini-1.9-8b has woven compact reasoning right into the answer itself:

• ⚡ Faster answers — no long hidden thought generation before the reply
• 🎯 Targeted reasoning — small reasoning steps are "interwoven" into the answer text
• 💾 Fewer tokens — the context window isn't wasted on large chains of thought

👁 Vision Engine: 1B → 2B

The model has an upgraded vision engine: the number of parameters was expanded from 1 billion to 2 billion compared to the previous version. The vision component is tightly integrated into the model itself (a unified mistral3 architecture), not a separate "bolted-on" module.

The practical results:

• 🚀 Fast image processing — processing a picture doesn't create a separate "bottleneck"; vision and language operate as a single whole
• 🖼 More accurate image understanding — the doubled vision engine is better at recognizing details, text within images (OCR), diagrams, and tables
• 💬 Multimodal dialogue — Image-Text-to-Text tasks: image + text question → text answer

Other Features

• Multilingual: support for dozens of languages, including Russian.
• Local execution: the GGUF format runs on CPU and GPU via Ollama, llama.cpp, vLLM, LM Studio, and other runtimes.
• Part of the Zubr lineup: the younger 3B version is alekringtonnn-ai/zubr-mini-1.9-3b (plus an MLX version for Mac).

Model Tree

mistralai/Ministral-3-8B-Base-2512
        │ (quantization / instruct version)
mistralai/Ministral-3-8B-Instruct-2512
        │
mistralai/Ministral-3-8B-Instruct-2512-GGUF
        │ (fine-tune by Zubr-AI)
Zubr-AI/zubr-mini-1.9-8b              ← GGUF version (THIS repository)

---

📊 Versions (Quantizations) in the Repository

The repository offers 6 GGUF quantization options:

💡 How to choose: the standard recommendation is Q4_0 (a quality/size balance). If you have enough memory, go with Q5_K_S: for a multimodal 8B model with a 2-billion-parameter vision engine, higher quantization gives a tangible boost to image recognition and reasoning quality. Q2/Q3 — only under strict memory constraints.

---

📥 How to Download

Requirements

• A CPU with AVX/AVX2 support (almost any modern x86) or an ARM device; a GPU is optional (NVIDIA/AMD accelerate inference; desirable for vision tasks).
• RAM: at least twice the size of the quantization file (see the table above); for Q4_0, a 12 GB RAM guideline.
• Free disk space: 3.1–6 GB depending on the version.

Method 1: Ollama (the easiest)

# Install Ollama: https://ollama.com
# The recommended version:
ollama run hf.co/Zubr-AI/zubr-mini-1.9-8b:Q4_0

# Other quantizations — just change the tag:
ollama run hf.co/Zubr-AI/zubr-mini-1.9-8b:Q5_K_S
ollama run hf.co/Zubr-AI/zubr-mini-1.9-8b:Q3_K_M

Ollama will automatically download the required GGUF file and start a chat in your terminal.

Method 2: llama.cpp

# Download and run in one command:
llama-cli -hf Zubr-AI/zubr-mini-1.9-8b:Q4_0

# Or download the file and specify it explicitly:
llama-cli -m zubr-mini-1.9-8b.Q4_0.gguf -p "Hello! Tell me about yourself."

Method 3: huggingface-cli (direct file downloads)

# Install the CLI
pip install -U "huggingface_hub[cli]"

# Download a specific quantization:
huggingface-cli download Zubr-AI/zubr-mini-1.9-8b \
  zubr-mini-1.9-8b.Q4_0.gguf

# Download the entire repository (all 6 quantizations, ~25 GB):
huggingface-cli download Zubr-AI/zubr-mini-1.9-8b

By default, files are saved to ~/.cache/huggingface/hub/.

Method 4: Direct links (browser / wget / curl)

curl -L -o zubr-mini-1.9-8b.Q4_0.gguf \
  "https://huggingface.co/Zubr-AI/zubr-mini-1.9-8b/resolve/main/zubr-mini-1.9-8b.Q4_0.gguf"

The same links also work in a browser: open the file page and click "Download", or append ?download=true to the URL.

Method 5: LM Studio / Jan (no command line)

1. Install LM Studio or Jan.
2. In the search tab, type zubr-mini-1.9-8b (or paste the repository link).
3. Select a quantization and click Download.
4. Start chatting in the built-in chat; images can be attached right in the interface.

---

🚀 How to Run

Via Ollama

ollama run hf.co/Zubr-AI/zubr-mini-1.9-8b:Q4_0

For multimodal requests (in the Ollama CLI), simply mention the image in the dialogue; in GUI clients, attach the image to your message.

Via llama.cpp (chat with an image)

llama-cli \
  -m zubr-mini-1.9-8b.Q4_0.gguf \
  --image photo.jpg \
  -p "What is in this image? List all the details."

Via vLLM (API server)

pip install vllm
vllm serve "Zubr-AI/zubr-mini-1.9-8b"

Once started, the model is available through an OpenAI-compatible API at http://localhost:8000/v1.

Via Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama.from_pretrained(
    repo_id="Zubr-AI/zubr-mini-1.9-8b",
    filename="*Q4_0.gguf",  # Or any other quantization
    verbose=False,
)

response = llm.create_chat_completion(
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello! Write a short poem about a bison."},
    ]
)
print(response["choices"][0]["message"]["content"])

---

⚙️ Recommended Generation Parameters

---

⚠️ Important Notes

• The model "thinks" right in the answer: the training principle of "context + question + the beginning of small reasoning steps" means that reasoning is part of generation. Don't disable intermediate steps in your runtime settings if the runtime marks them.
• A 2-billion-parameter vision engine: it is integrated into the shared weights, so high-quality quantizations (Q4_0/Q5_K_S) preserve image recognition quality much better than low ones (Q2/Q3) — the vision part suffers from heavy quantization first.
• Memory: the 8B model requires more RAM than the 3B version: aim for 12+ GB for comfortable Q4_0 operation, 16 GB for Q5_K_S.
• Hallucinations: like any LLM, the model may generate factually incorrect information — verify important facts.
• Small quantizations (Q2/Q3): for a multimodal model, the quality loss is more noticeable than for a text-only 3B one — use them only under strict memory constraints.
• The repository is fresh: at the time of writing this documentation, the repository has no README or configuration files — check the model page for licensing and details.

---

🔗 Links

• Repository (GGUF): https://huggingface.co/Zubr-AI/zubr-mini-1.9-8b
• Younger 3B version: https://huggingface.co/alekringtonnn-ai/zubr-mini-1.9-3b
• MLX version of the 3B for Apple Silicon: https://huggingface.co/alekringtonnn-ai/zubr-mini-1.9-3b-mlx
• Zubr Audio model: https://huggingface.co/alekringtonnn-ai/zubr-audio
• Ollama: https://ollama.com
• llama.cpp: https://github.com/ggml-org/llama.cpp
• vLLM: https://github.com/vllm-project/vllm
• LM Studio: https://lmstudio.ai

---

Author: Zubr-AI (alekringtonnn-ai) · Base model: Mistral 3 8B (Ministral-3-8B) · Type: multimodal LLM (Image-Text-to-Text) · Format: GGUF (all platforms)