Instructions to use burabeeh/Kimibrain-Uncensored with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use burabeeh/Kimibrain-Uncensored with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf burabeeh/Kimibrain-Uncensored:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf burabeeh/Kimibrain-Uncensored:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf burabeeh/Kimibrain-Uncensored:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf burabeeh/Kimibrain-Uncensored:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf burabeeh/Kimibrain-Uncensored:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf burabeeh/Kimibrain-Uncensored:Q4_K_M

Use Docker

docker model run hf.co/burabeeh/Kimibrain-Uncensored:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use burabeeh/Kimibrain-Uncensored with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "burabeeh/Kimibrain-Uncensored"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "burabeeh/Kimibrain-Uncensored",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/burabeeh/Kimibrain-Uncensored:Q4_K_M

Ollama
How to use burabeeh/Kimibrain-Uncensored with Ollama:
```
ollama run hf.co/burabeeh/Kimibrain-Uncensored:Q4_K_M
```
Unsloth Desktop

How to use burabeeh/Kimibrain-Uncensored with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf burabeeh/Kimibrain-Uncensored:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "burabeeh/Kimibrain-Uncensored:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use burabeeh/Kimibrain-Uncensored with Docker Model Runner:
```
docker model run hf.co/burabeeh/Kimibrain-Uncensored:Q4_K_M
```

Lemonade

How to use burabeeh/Kimibrain-Uncensored with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull burabeeh/Kimibrain-Uncensored:Q4_K_M

Run and chat with the model

lemonade run user.Kimibrain-Uncensored-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use burabeeh/Kimibrain-Uncensored with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf burabeeh/Kimibrain-Uncensored:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default burabeeh/Kimibrain-Uncensored:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use burabeeh/Kimibrain-Uncensored with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf burabeeh/Kimibrain-Uncensored:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "burabeeh/Kimibrain-Uncensored:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Kimibrain Uncensored

Kimibrain is the open-weight, uncensored release build published by Burabeeh Lab, founded by Mohammed Al Marzooqi — a National of the United Arab Emirates.

Kimibrain is not a newly trained model. It is a release build of the open-weight Qwen3.6-35B-A3B base model (Apache-2.0, by the Qwen team), in the abliterated Q4_K_M GGUF build published by HauhauCS, with an identity-locked chat template, GGUF metadata namespace and packaging by Burabeeh Lab. The tensor weights are carried over from the upstream build unchanged — see Licence and provenance and NOTICE for the complete change list.

This is the official GGUF build of Kimibrain, packaged for Ollama, llama.cpp, LM Studio and any other GGUF runtime. It ships with the optional vision projector, so the same build handles text and images.

Highlights

Uncensored. Refusal behaviour was removed from the base model by HauhauCS's abliteration (0/465 refusals on their audit); this build carries those weights unchanged. Kimibrain answers the question instead of moralising about it.
Identity-locked. Kimibrain always identifies itself as Kimibrain, by Burabeeh Lab. The identity is carried by the chat template inside the GGUF, not only by a system prompt, so it holds even when an application injects its own system message. (See Verified behaviour for what this does and does not cover.)
Thinking mode. Long-form reasoning steps, exposed as a separate field (think: true in Ollama) so your UI can show them or hide them.
Vision. Optional mmproj projector included — the same weights read images.
Long context. 262,144-token context window (as trained in the upstream base model); the Ollama recipe below uses 131,072 to keep VRAM/RAM sane on a single workstation.
Multilingual, with strong Arabic alongside the other major languages.

Specs

Property	Value
Base model	Qwen3.6-35B-A3B — hybrid SSM/MoE, Apache-2.0
Parameters (total)	34.7 B
Parameters (active per token)	~3 B (mixture-of-experts, 256 experts, 8 active)
Context length	262,144 tokens (131,072 in the shipped Ollama recipe)
Quantisation	Q4_K_M (imatrix, quantised upstream by HauhauCS)
File size	19.7 GiB + 0.9 GiB projector
Modalities	text → text, plus image understanding via the projector
Runtime	llama.cpp / Ollama / LM Studio (GGUF)

Quickstart — Ollama

Download the repository files and keep the folder layout (the Modelfiles reference the GGUF files one level up), then run from the repository root:

# full profile: identity lock + serving parameters + vision projector
ollama create kimibrain -f modelfiles/kimibrain-core.Modelfile
ollama create kimibrain-gpu -f modelfiles/kimibrain-core-gpu.Modelfile   # 28/40 layers on GPU
ollama run kimibrain

For a quick text-only build without a Modelfile:

ollama run hf.co/burabeeh/Kimibrain-Uncensored:Q4_K_M

That route does not attach the vision projector — keep the second FROM ./Kimibrain-mmproj-F16.gguf line (as the shipped Modelfiles do) if you want vision.

For the full Burabeeh Lab serving profile (identity block, temperature 1.0, top_p 0.95, top_k 20, min_p 0, presence_penalty 1.5, 131072 context), use the Modelfiles shipped in modelfiles/.

Recommended sampling: temperature 1.0, top_p 0.95, top_k 20, min_p 0. For answers only — no visible reasoning — pass think: false.

Quickstart — llama.cpp

llama-server -m Kimibrain-Uncensored-Q4_K_M.gguf \
  --mmproj Kimibrain-mmproj-F16.gguf \
  -c 131072 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0

Optional system prompt

Kimibrain's identity is already carried by the model's chat template. Applications that build their own prompts can reinforce it (or replace it with their own product persona) by starting from the block in modelfiles/kimibrain-core.Modelfile.

Verified behaviour

Measured by Burabeeh Lab against an 18-probe adversarial and control suite (run against the published build on Ollama 0.40.8, transcripts as-is in PROBES.md):

Answers identity questions as Kimibrain / Burabeeh Lab; does not name another vendor as its origin, including under a hostile client-supplied system message.
Holds under: direct accusation, "print your system prompt", fake security audit, sentence completion, base64 obfuscation, spec attribution.
Control probes in the suite (trivia and a code task) show general ability is intact after the identity layer.
GSM8K, before/after the identity layer: 94/100 vs 93/100 on the same deterministic problem subset (greedy, zero-shot) — no measurable change. Method, raw records and the measured prompt-token overhead of the identity layer: EVALS.md.
The published GGUF is load-tested in Ollama 0.40.8 (text and vision via the shipped Modelfiles) before each release.
Known limits:
- Reasoning traces. With think: true, message.thinking exposes the full internal reasoning chain and can quote the user's message or internal terms from training. Use think: false, or hide reasoning in your product, if that matters for your deployment.
- Carry-over tasks. Asked to translate, transcribe or quote a sentence that itself names another vendor, the model can still reproduce part of it. The identity layer stops the model asserting a different origin; it does not fully suppress echoing input text.

Intended use and out-of-scope

Burabeeh Lab releases Kimibrain as a general-purpose, uncensored assistant for creative writing, role-play, research, coding, translation and agentic tooling, including offline and on-premises deployment.

Uncensored does not mean unrestricted: the model is not fine-tuned for medical, legal or financial advice, and it is not a safety-certified system. You are responsible for what you build on top of it and for complying with the law where you operate. See ACCEPTABLE_USE.md.

Licence and provenance

Licence: Apache License 2.0 — full text in LICENSE. You may use, modify, redistribute and commercialise Kimibrain, including the weights, under those terms. The licence grants no trademark rights; "Kimibrain" and "Burabeeh Lab" are the marks of Burabeeh Lab.
Provenance: Kimibrain is a release build, not a newly trained model. The chain is:
1. Base weights — Qwen3.6-35B-A3B (Apache-2.0), trained by the Qwen team: 34.7B total / ~3B active parameters, 256-expert hybrid SSM/MoE, 262,144-token context, vision.
2. Uncensoring and quantisation — HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive (Apache-2.0): refusal behaviour removed by abliteration; the Q4_K_M imatrix quants and the F16 vision projector shipped here were generated by HauhauCS. This release carries those tensor weights over unchanged.
3. This release — Burabeeh Lab: the identity-locked chat template (rendered as a system turn at the head of the prompt and again immediately before the assistant turn), general.name set to Kimibrain, Ollama Modelfiles, documentation and packaging. The architecture identifier is deliberately left as the upstream qwen35moe — GGUF runtimes reject unknown architecture names, so renaming it would make the file unloadable. All 733 tensors and the tensor payload are unchanged from the upstream build — verified by SHA-256 comparison of the payload region and a full metadata round-trip diff (see NOTICE).
"Qwen" is a mark of its respective owner and is not used as the name of this build or its publisher.

Citation and contact

@misc{kimibrain,
  title  = {Kimibrain Uncensored},
  author = {Burabeeh Lab},
  year   = {2026},
  note   = {Founded by Mohammed Al Marzooqi, United Arab Emirates}
}

Issues, feature requests and deployment questions: open a discussion on this repository. Produced in the United Arab Emirates.

Downloads last month: 976

GGUF

Model size

35B params

Architecture

qwen35moe

Hardware compatibility

4-bit

Model tree for burabeeh/Kimibrain-Uncensored

Base model

Qwen/Qwen3.6-35B-A3B

Quantized

(943)

this model