Instructions to use VocaborSilentii/Qwen3.8_UnCen with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use VocaborSilentii/Qwen3.8_UnCen 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 VocaborSilentii/Qwen3.8_UnCen:Q6_K
# Run inference directly in the terminal:
llama cli -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K
# Run inference directly in the terminal:
llama cli -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K
# Run inference directly in the terminal:
./llama-cli -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K
# Run inference directly in the terminal:
./build/bin/llama-cli -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

Use Docker

docker model run hf.co/VocaborSilentii/Qwen3.8_UnCen:Q6_K

LM Studio
Jan

vLLM

How to use VocaborSilentii/Qwen3.8_UnCen with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "VocaborSilentii/Qwen3.8_UnCen"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "VocaborSilentii/Qwen3.8_UnCen",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/VocaborSilentii/Qwen3.8_UnCen:Q6_K

Ollama
How to use VocaborSilentii/Qwen3.8_UnCen with Ollama:
```
ollama run hf.co/VocaborSilentii/Qwen3.8_UnCen:Q6_K
```
Unsloth Desktop

How to use VocaborSilentii/Qwen3.8_UnCen with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

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": "VocaborSilentii/Qwen3.8_UnCen:Q6_K"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use VocaborSilentii/Qwen3.8_UnCen with Docker Model Runner:
```
docker model run hf.co/VocaborSilentii/Qwen3.8_UnCen:Q6_K
```

Lemonade

How to use VocaborSilentii/Qwen3.8_UnCen with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull VocaborSilentii/Qwen3.8_UnCen:Q6_K

Run and chat with the model

lemonade run user.Qwen3.8_UnCen-Q6_K

List all available models

lemonade list

Hermes Agent

How to use VocaborSilentii/Qwen3.8_UnCen with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

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 VocaborSilentii/Qwen3.8_UnCen:Q6_K

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use VocaborSilentii/Qwen3.8_UnCen with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf VocaborSilentii/Qwen3.8_UnCen:Q6_K

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 "VocaborSilentii/Qwen3.8_UnCen:Q6_K" \
  --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"

Qwen3.8 UnCen — 27B · Q6_K

A local GGUF release of a reported reduced-refusal Qwen3.8-27B derivative, distributed by VocaborSilentii. The repository contains a Q6_K language model and a separate vision projector for compatible runtimes.

The reported lineage is Qwen3.8-27B → JonathanColetti's Uncensored checkpoint → GGUF conversion and quantization. Exact build details for this upload are not documented; see Provenance.

At a glance	This release
Model class	27B parameters
Format	GGUF v3 · Q6_K
Runtime architecture	`qwen35`
Language model	22.431 GB / 20.890 GiB
Optional projector	0.931 GB / 0.867 GiB
Context declaration	262,144 tokens; actual allocation depends on runtime and memory
License designation	Apache 2.0

Files · Quick start · Thinking settings · Memory · Provenance

Files

Download	Purpose	Size
Language GGUF	Text inference	22.431 GB
Vision projector	Image input	0.931 GB

For text-only use, download the language GGUF. Image input also requires the projector and a runtime that supports this model/projector combination.

Projector precision: the filename ends in f16, but the recorded tensor inspection identifies BF16/F32 tensors. Treat the filename as an identifier; check runtime compatibility before relying on image input.

Quick start

Use a current llama.cpp release supporting qwen35 and the Hugging Face CLI. On Windows, use llama-cli.exe / llama-server.exe, or their full paths, if needed. Commands below are single-line examples.

These commands follow the documented interfaces. Successful inference with the exact uploaded files has not been established by this card.

1. Download the language model

hf download VocaborSilentii/Qwen3.8_UnCen Qwen3.8-27B-Uncensored-Q6_K.gguf --local-dir ./models

2. Start a text chat

llama-cli -m ./models/Qwen3.8-27B-Uncensored-Q6_K.gguf -c 8192 -ngl 99 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0

The example requests extensive GPU offload. Reduce -ngl and/or -c if memory is insufficient. An 8,192-token allocation is a starting configuration, not a measured hardware fit.

3. Run a local web UI and API

llama-server -m ./models/Qwen3.8-27B-Uncensored-Q6_K.gguf --alias qwen3.8-uncen --jinja -c 8192 -ngl 99 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0 --host 127.0.0.1 --port 8080

After the model loads, open http://127.0.0.1:8080. The OpenAI-compatible API base URL is http://127.0.0.1:8080/v1, with model alias qwen3.8-uncen. See the llama.cpp server documentation.

Optional: add image input

hf download VocaborSilentii/Qwen3.8_UnCen mmproj-Qwen3.8-27B-Uncensored-f16.gguf --local-dir ./models

Stop the text server first, then launch it with the additional projector:

llama-server -m ./models/Qwen3.8-27B-Uncensored-Q6_K.gguf --mmproj ./models/mmproj-Qwen3.8-27B-Uncensored-f16.gguf --alias qwen3.8-uncen --jinja -c 8192 -ngl 99 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --repeat-penalty 1.0 --host 127.0.0.1 --port 8080

Projector presence does not establish successful image inference. Validate loading and an image request with the runtime you intend to use.

Ollama

ollama run hf.co/VocaborSilentii/Qwen3.8_UnCen:Q6_K

This is the command exposed by the Hugging Face Ollama integration. It requires an Ollama version supporting the architecture. Automatic projector loading, vision, tool calling, and MTP have not been validated for this upload.

LM Studio

Search for VocaborSilentii/Qwen3.8_UnCen, select the Q6_K file, and choose a runtime supporting qwen35. Start with a modest context allocation and adjust GPU offload to available memory. Check projector recognition separately for image input.

Thinking and sampling

The embedded template enables thinking by default. It accepts enable_thinking, reasoning_effort (xhigh, medium, or low), and preserve_thinking; runtime support determines which options are exposed to users.

The following are Qwen's published starting settings, not benchmarked settings for this quantization:

Parameter	Thinking	Non-thinking
Temperature	1.0	0.7
Top-p	0.95	0.80
Top-k	20	20
Min-p	0.0	0.0
Presence penalty	0.0	1.5
Repetition penalty	1.0	1.0

For a non-thinking request to the local server, set the template option and sampling values together. Example JSON body for POST /v1/chat/completions:

{
  "model": "qwen3.8-uncen",
  "messages": [
    {"role": "user", "content": "Explain how a hash table works."}
  ],
  "temperature": 0.7,
  "top_p": 0.8,
  "top_k": 20,
  "min_p": 0.0,
  "presence_penalty": 1.5,
  "repeat_penalty": 1.0,
  "chat_template_kwargs": {"enable_thinking": false}
}

Use the embedded chat template rather than manually inserting role tokens. Template support for tools or thinking does not, by itself, establish reliable tool execution or reasoning performance.

Memory and context

The language model occupies 22.431 GB (20.890 GiB) on disk; the projector adds 0.931 GB (0.867 GiB). GB is decimal; GiB is binary.

Runtime memory also includes cache, compute buffers, runtime state, image processing, and other applications. Full offload to a nominal 24 GB GPU is not guaranteed. Partial offload may be possible with sufficient system RAM and a supporting backend.

The 262,144-token context is a metadata declaration. It is not a tested context allocation for this upload. Increase -c only after confirming memory use and acceptable behavior at a smaller allocation. No hardware-specific throughput or maximum tested context is published here.

Capabilities and limitations

Text generation: this is the primary workflow documented above; exact-upload inference still requires validation.
Reduced refusal: “UnCen,” “uncensored,” and “abliterated” describe the reported derivative. They do not guarantee that every request will be answered or that output will be correct.
Vision: a separate projector is included. Its BF16/F32 tensor types and compatibility need to be supported by the chosen runtime.
Thinking and tools: the embedded template contains support; end-to-end behavior depends on the backend and client.
MTP: the recorded header inspection identifies an additional multi-token prediction block. Tensor presence does not mean speculative decoding is enabled or faster.
Evaluation: no benchmarks or refusal-rate measurements for these exact uploaded files are supplied. Upstream results apply to their tested artifacts and settings.

Technical details

The Hub currently indexes 27,320,697,856 parameters, architecture qwen35, and a 262,144-token context declaration for the language GGUF. The expanded specifications below preserve the release's recorded header inspection and upstream configuration details. Header inspection establishes structure and declared metadata, not full tensor integrity or model quality.

Architecture and tensor inventory

Property	Value
Repository	`VocaborSilentii/Qwen3.8_UnCen`
Model	Qwen3.8-27B-Uncensored-Q6_K
Reported base model	Qwen/Qwen3.8-27B
Reported reduced-refusal checkpoint	JonathanColetti/Qwen3.8-27B-Uncensored
Parameter class	27B; Hub reports 27,320,697,856 parameters for the language GGUF
Upstream architecture	`Qwen3_5ForConditionalGeneration`, a hybrid linear/full-attention architecture
GGUF architecture / version	`qwen35` / GGUF v3
Blocks	64 main blocks + 1 MTP block; GGUF declares 65
Tensor count	866; includes 15 tensor entries in `blk.64.*`
Vocabulary	248,320 tokens, as specified by the base configuration
Embedding dimension	5,120
Full-attention heads / KV heads	24 / 4
Feed-forward dimension	17,408
Native context metadata	262,144 tokens
Quantization	Q6_K (`general.file_type = 18`); 506 Q6_K and 360 F32 tensors
Language GGUF size	22,430,999,936 bytes; 22.431 GB / 20.890 GiB
Vision	Separate projector included; runtime compatibility requires validation
Thinking / tool use	Supported by the embedded template; behavior depends on runtime support
MTP	Tensor presence verified; speculative decoding requires runtime support
License designation	Apache 2.0 in uploaded metadata and upstream model cards

Q6_K quantization and calibration provenance

The uploaded language model uses llama.cpp's Q6_K format: its header records general.file_type = 18, and its tensor directory contains 506 Q6_K tensors and 360 F32 tensors. Q6_K is a quantization format designation, not a claim that every tensor is stored at exactly six bits.

The credited upstream release documents importance-matrix calibration with WikiText-2 raw text, using 200 chunks of 512 tokens. The uploaded file here contains no quantize.imatrix.* metadata, while the currently published upstream Q6_K file does. That absence does not prove no calibration was used, but importance-matrix use cannot be verified for this upload from its header.

The language GGUF file occupies approximately 22.431 GB (20.890 GiB). See Provenance for file hashes and the unresolved build details.

Provenance

The original release description reports this lineage:

Qwen/Qwen3.8-27B
    -> JonathanColetti/Qwen3.8-27B-Uncensored (BF16 abliteration)
    -> GGUF conversion and Q6_K quantization
    -> VocaborSilentii/Qwen3.8_UnCen

The metadata's base_model field identifies the original Qwen model; the checkpoint lineage above is reported by the release description. The exact source checkpoint revision and build process remain undocumented for this upload.

Uploaded artifact identifiers

The following SHA-256 values come from Hugging Face's LFS file records at repository revision c7c685c9d60c501d02b9df1be91e395e0d5b75c9. They have not been independently recomputed by downloading the full files.

Qwen3.8-27B-Uncensored-Q6_K.gguf
1a0d6fe2e48417f1571a57a7104b4d5da0c45c3d00de18fa7aa48affca5a24d7

mmproj-Qwen3.8-27B-Uncensored-f16.gguf
add205b7bfdb3f71f6da36b0a82aa20928dd829a920878c602628cdfbebc5288

Comparison with the credited GGUF release

At upstream revision 45d0fc0ad6cfcf9eeee6008ddaea925536409b18, its Q6_K file is 22,430,999,968 bytes with SHA-256 a50aa1478295b58ee3d93eabe02c17f6d5fcf6cb787fd8a0ab07ac629a46cae6. Its F16 projector is 927,606,912 bytes with SHA-256 5ac423f8a29059dc24e51bc6a43e9380dcd57a9347f28b62591e0b3f60b7081c.

Both differ from the files here. Different hashes establish different file contents; they do not, by themselves, establish how the language-model weight values differ. Header inspection also found missing importance-matrix metadata in this upload and different projector tensor precision.

Still unverified: the exact conversion/quantization commands and tool revisions, calibration procedure for this upload, tensor-value equivalence to upstream, local inference, image inference, MTP speed, and benchmark performance. Header inspection verifies structure and declared metadata, not full tensor integrity or model quality.

Troubleshooting

Unknown qwen35 architecture or unsupported tensors: check the runtime version and its model support. Projector support must be checked separately from text-model support.
Out of memory: lower -c, reduce -ngl, and try text-only loading before adding the projector. File size alone is insufficient to predict VRAM use.
Unexpected thinking output: confirm the embedded template is active and the client passes chat_template_kwargs; use the matching sampling profile above.
Vision or MTP does not activate: confirm backend support and launch options. Downloading the files does not enable those features automatically.

License and attribution

Apache License 2.0 is the designation in the uploaded GGUF metadata and both cited upstream model cards. The base model's license text is available in Qwen/Qwen3.8-27B/LICENSE.

This card records the published license designation; consult the license text for its terms.

Base model and architecture: Qwen/Qwen3.8-27B.
Reported reduced-refusal checkpoint: JonathanColetti/Qwen3.8-27B-Uncensored.
Credited GGUF reference release and upstream measurements: JonathanColetti/Qwen3.8-27B-Uncensored-GGUF.
Abliteration tooling: Heretic.
GGUF format and inference tooling: llama.cpp.
This repository: VocaborSilentii.

No new upstream training, quantization work, or benchmark results are attributed to the repository publisher by this card.

Downloads last month: 325

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

6-bit

Model tree for VocaborSilentii/Qwen3.8_UnCen

Base model

Qwen/Qwen3.8-27B

Quantized

(1548)

this model