Instructions to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF 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 lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S
# Run inference directly in the terminal:
llama cli -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S
# Run inference directly in the terminal:
llama cli -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

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 lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S
# Run inference directly in the terminal:
./build/bin/llama cli -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

Use Docker

docker model run hf.co/lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

Llama new
LM Studio
Jan

vLLM

How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

Ollama
How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with Ollama:
```
ollama run hf.co/lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S
```
Unsloth Desktop

How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

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": "lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with Docker Model Runner:
```
docker model run hf.co/lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S
```

Lemonade

How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

Run and chat with the model

lemonade run user.Holo4-27B-Abliterix-GSQ-RCO-GGUF-IQ2_S

List all available models

lemonade list

Hermes Agent

How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

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 lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S

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 "lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF:IQ2_S" \
  --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"

Holo4-27B — Abliterated, GSQ-RCO-style GGUF

Per-tensor mixed-precision GGUF builds of the best-fidelity public uncensor of Holo4-27B, made for llama.cpp.

What this model is

Hcompany/Holo4-27B is a 27B vision-language model for Computer Use — the strongest Holo4, at OSWorld 85.2% at $0.08/task and OSWorld 2.0 61.7% at $1.22/task. Those are vendor BF16 numbers, not measurements of these quants.

The vendor calls the architecture "Qwen3.8 dense" (built on Qwen/Qwen3.8-27B) — "dense" means no routed experts, not a plain transformer. It is a hybrid: 64 blocks, 16 with full attention and 48 with linear attention (GDN: attn_qkv, attn_gate, ssm_out, ssm_conv1d, ssm_alpha/beta), dense MLPs everywhere (ffn_gate/up/down, 17,408 hidden), 248,320-token vocab, 262,144-token context.

License is non-commercial CC BY-NC 4.0 — the base Qwen3.8-27B is Apache-2.0 and both licenses ship alongside. Fine to run locally; not commercially servable.

The ablation ("uncensor")

No Heretic pass exists for this architecture — Heretic's activation-based refusal-direction search has not been run on Qwen3.8-hybrid weights as of writing. The uncensor used here is timteh673/Holo4-27B-Abliterix, a directional ablation (abliteration: projecting the refusal direction out of activations), published at BF16 with ~54 GB of safetensors.

Files

File	Size	BPW	Ladder shape
`quant/Holo4-27B-Abliterix-GSQ-RCO-IQ2_XXS.gguf`	7.28 GiB	2.33	64 blocks @ IQ2_XXS; embeddings Q4_K
`quant/Holo4-27B-Abliterix-GSQ-RCO-IQ2_S.gguf`	7.83 GiB	2.50	25 edge @ IQ2_S, 39 inner @ IQ2_XXS; embeddings Q4_K
`quant/Holo4-27B-Abliterix-GSQ-RCO-IQ3_XXS.gguf`	9.35 GiB	2.99	10 edge @ IQ3_XXS, 39 mid @ IQ2_S, 15 inner @ IQ2_XS
`quant/Holo4-27B-Abliterix-GSQ-RCO-IQ3_S.gguf`	10.96 GiB	3.50	19 edge @ IQ3_S, 16 mid @ IQ3_XXS, 29 inner @ IQ2_S
`quant/mtp-Holo4-27B-Abliterix-GSQ-RCO-Q8_0.gguf`	441 MiB	—	MTP draft head (Qwen3.8-27B base) for speculative decoding

Only the final quantized builds and the MTP draft head are hosted here. The intermediate Q6_K source, the importance matrix, and the vision projector are linked below instead of being mirrored in this repository.

Source and provenance

Hcompany BF16 master
   → timteh673/Holo4-27B-Abliterix (BF16 safetensors, no quant loss)
   → BF16 GGUF conversion (2.3 / 2.5 tiers, no intermediate quantization)
   → mradermacher i1 Q6_K quant (21,049 MiB, 6.57 bpw; 3.0 / 3.5 tiers)

The 2.3 / 2.5 bpw tiers are first-generation builds from the BF16 uncensor master. The 3.0 / 3.5 tiers sit one Q6_K generation above it. No abliteration was redone for either path.

BF16 uncensored master: timteh673/Holo4-27B-Abliterix
Q6_K source: mradermacher/Holo4-27B-Abliterix-i1-GGUF → Holo4-27B-Abliterix.i1-Q6_K.gguf
Importance matrix: mradermacher/Holo4-27B-Abliterix-i1-GGUF → Holo4-27B-Abliterix.imatrix.gguf — built against the same Abliterix weights. Coverage: all 496 significant 2-D matrices; token_embd.weight / output.weight carry no rows but are pinned high and do not need them.
Vision projector: mradermacher/Holo4-27B-Abliterix-GGUF → Holo4-27B-Abliterix.mmproj-f16.gguf — BF16, required for image input.

Quantization

"GSQ-RCO-style" means a format-compatible equivalent of the GSQ/RCO 3-bit recipe, not a bit-exact reproduction. The published GSQ/RCO engines emit only HF compressed-tensors, have no GGUF export, and target sm_90+ GPUs.

The allocation spans every 2-D matrix of a block: dense MLPs, attention, and the GDN ssm_out projection. Edge blocks (by distance from block 0/63) take the top rung; redundant inner blocks take the bottom rung.

Fixed tensors:

3.50 tier: embeddings pinned Q8_0 (9.4% of parameters on the 248k vocab).
3.00 tier: embeddings dropped to Q6_K — with no expert mass to flex, the only way to reach 3.00 overall is to spend bits across everything; Q8 embeddings alone would eat 26% of the budget.
2.50 / 2.30 tiers: embeddings at Q4_K; the budget left for the block ladder is too small for Q6_K at these sizes. The 2.50 tier puts 25 edge blocks at IQ2_S and the 39 inner blocks at IQ2_XXS; the 2.30 tier runs all 64 blocks at IQ2_XXS.
1-D norms/biases fp32.
Vision projector kept at vendor BF16.

These are non-uniform files — no single quant type describes them, which is why the HF quant badge is left blank rather than filled with something wrong.

Usage

llama-cli -m quant/Holo4-27B-Abliterix-GSQ-RCO-IQ3_S.gguf \
  -p "Explain mixed-precision quantization." -ngl 99

# vision
llama-mtmd-cli -m quant/Holo4-27B-Abliterix-GSQ-RCO-IQ3_S.gguf \
  --mmproj Holo4-27B-Abliterix.mmproj-f16.gguf \
  --image screenshot.png -p "Describe this screen."

All four builds were smoke-tested: load and coherent output with CPU-only llama.cpp f3f31a7 tooling. Raw CPU generation is ~1.4–1.9 t/s — the GDN/SSM ops are slow without GPU kernels; that is a smoke signal, not a serving number. The 2-bit files are aggressive by design; use IQ3 for quality-sensitive work.

Speculative decoding (MTP)

The vendor Holo checkpoint declares an MTP head but ships no mtp.* weights, so these files have no usable built-in MTP layer. The draft head here is the MTP block from the matching base model (Qwen/Qwen3.8-27B), quantized to Q8_0. Tokenizer and dimensions match the Holo weights exactly, and Q8_0 acceptance was identical to BF16 in testing.

llama-server -m quant/Holo4-27B-Abliterix-GSQ-RCO-IQ3_S.gguf \
  -md quant/mtp-Holo4-27B-Abliterix-GSQ-RCO-Q8_0.gguf \
  --spec-type draft-mtp --spec-draft-n-max 3 -ngl 99

Measured greedy CPU acceptance on a code prompt: 74.1% (mean accepted length 3.15). The draft file has no token_embd.weight of its own; load it with -md, not as a standalone model. A borrow_shared_tensor log line is expected and harmless.

Honest limits

The IQ3 builds sit one quant generation above the BF16 Abliterix master; the IQ2 builds are first-generation from BF16. That master itself is the only public uncensor of this architecture and predates any Heretic work; if a Heretic-for-Qwen3.8 pass ever lands, re-quant from that.
Abliteration on hybrid linear-attention weights has less community validation than on plain transformers. Refusal removal is behavioural, and edge cases can refuse or misbehave under adversarial prompting.
Non-commercial license — do not serve this commercially.

Downloads last month: 2,175

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

2-bit

3-bit

8-bit

Model tree for lavendercomputing/Holo4-27B-Abliterix-GSQ-RCO-GGUF

Base model

Qwen/Qwen3.8-27B

Finetuned

Hcompany/Holo4-27B

Quantized

(14)

this model