---

🧠 Overview

DanyAI-MiaBuild-9b-Uncensored is not a typical unlocked model.

It is the first official output of the MiaBuild™ system — a proprietary, closed-source deterministic build system primarily designed for critical infrastructure and industrial software. The model was processed through the MiaBuild pipeline, which internally performs the entire transformation process — fine-tuning, abliteration, system prompt replacement, and multi-stage validation — delivering a cryptographically signed, audit-verified artifact.

---

🏭 What is MiaBuild™?

MiaBuild is 100% independent of LLMs. It is an offline, deterministic project factory.

MiaBuild is a proprietary two-part build system developed at DanyAI. It does not require any external language model to operate — it generates complete, provable, production-ready software products from formalized industrial knowledge without nondeterministic elements.

⚙️ Key pipeline features

🏆 Industry standards

• DO-178C — Avionics software (Level A)
• MIL-STD-882E — Military system safety
• IEC 61508 — Functional safety (SIL 4)
• ISO 26262 — Automotive safety (ASIL D)

MiaBuild does not disclose its internal architecture. The pipeline is proprietary and closed.

---

📊 Determinism: two separate metrics

🏭 MiaBuild™ pipeline: 100% bit-for-bit

MiaBuild itself operates fully deterministically. The same input always produces a bit-for-bit identical output — no exceptions, no variance. This is a core architectural property of the system, enforced at the tooling level and verified through independent CI controls.

🤖 Processed model: 93.78%

When MiaBuild processes a language model, it reaches up to 93.78% determinism — measured using the MiaBuild 50-run test, where the same prompt is executed 50 times and outputs are compared token-by-token.

This is an exceptional result because LLMs are inherently nondeterministic systems. Floating-point GPU operations, different CUDA versions, thread scheduling, and processor thermal management always introduce some variance — even with temperature=0. This is a physical limitation, not an implementation flaw.

Mainstream models do not even measure this:

The remaining ~6.22% represents the physical limit of floating-point arithmetic at the hardware level — a limit that cannot be eliminated without full control over the hardware stack.

Why this matters for critical systems

• 🏭 Industrial reproducibility — the same input tomorrow produces the same output as today
• ⚖️ Legal traceability — regulated industries must reproduce the exact output used as the basis for a decision
• 🧪 Reliable testing — without determinism there are no reliable automated tests for LLM outputs
• 🤖 Multi-agent stability — nondeterminism in one agent propagates exponentially through agent chains

MiaBuild pipeline = 100% bit-for-bit. Model = 93.78% — which is close to the physical limit of what is achievable with LLMs.

---

🔧 Technical specifications

---

🌍 Language support

Primary languages (highest quality):

🇸🇮 Slovenian  ·  🇷🇸 Serbian  ·  🇭🇷 Croatian  ·  🇬🇧 English

Additionally supports 197+ languages from the Qwen3.5-9B base model training.

---

📥 Downloads (GGUF)

💼 Q6_K, Q8_0, and BF16 are available exclusively for paying customers. For access contact: 📧 cvijetic.danijel@gmail.com

💡 Multimodal usage: Load the main GGUF file and the mmproj file together.

---

⚙️ Recommended settings

🧩 Thinking mode — recommended for reasoning and analysis

temperature=0.6  |  top_p=0.95  |  top_k=20  |  min_p=0

🎨 Creative / non-thinking mode

temperature=0.7  |  top_p=0.8  |  top_k=20  |  min_p=0

⚠️ Important: Use at least 128K context. The model performs significantly better with generous context allocation.

---

🚀 Usage

Supported runtimes: llama.cpp · LM Studio · Jan.ai · koboldcpp · Ollama · vLLM · SGLang · KTransformers

llama.cpp — text only

./llama-cli \
  -m DanyAI-MiaBuild-9b-Q6_K.gguf \
  -c 131072 \
  --temp 0.6 \
  -ngl 99

llama.cpp — with vision support

./llama-cli \
  -m DanyAI-MiaBuild-9b-Q6_K.gguf \
  --mmproj mmproj-DanyAI-MiaBuild-9b-BF16.gguf \
  -c 131072 \
  --temp 0.6 \
  -ngl 99

Ollama

ollama run danyai-miabuild-9b

---

💡 Model capabilities

• 🔧 Tool calling and function calling — native support
• 🤖 Multi-agent workflows — suitable as the backbone of agent systems
• 📄 Long document analysis — up to 262K native context
• 👁️ Vision / image input — load the mmproj file together with the main model
• 💻 Code generation — full preservation of base model capabilities
• 🌍 Multilingual support — first-class support for South Slavic languages

---

⚠️ Important notes

This model is fully unlocked and does not refuse prompts. Suitable for:

• 🏭 Critical infrastructure and industrial systems where refusals are undesirable
• 🤖 Custom agent systems where refusals would break functionality
• 🏢 Enterprise deployments requiring unrestricted model behavior
• 🔞 Adult content platforms (where legally permitted)
• 🔬 Research and development

You are solely responsible for generated content. Ensure compliance with applicable laws in your jurisdiction.

---

📬 Services and contact

Available throughout the EU and the Balkans.

📧 cvijetic.danijel@gmail.com

---