Instructions to use Cyronius/Qwen3.6-27B-A2.8B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use Cyronius/Qwen3.6-27B-A2.8B 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 Cyronius/Qwen3.6-27B-A2.8B
# Run inference directly in the terminal:
llama cli -hf Cyronius/Qwen3.6-27B-A2.8B

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Cyronius/Qwen3.6-27B-A2.8B
# Run inference directly in the terminal:
llama cli -hf Cyronius/Qwen3.6-27B-A2.8B

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 Cyronius/Qwen3.6-27B-A2.8B
# Run inference directly in the terminal:
./llama-cli -hf Cyronius/Qwen3.6-27B-A2.8B

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 Cyronius/Qwen3.6-27B-A2.8B
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Cyronius/Qwen3.6-27B-A2.8B

Use Docker

docker model run hf.co/Cyronius/Qwen3.6-27B-A2.8B

LM Studio
Jan

vLLM

How to use Cyronius/Qwen3.6-27B-A2.8B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Cyronius/Qwen3.6-27B-A2.8B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Cyronius/Qwen3.6-27B-A2.8B",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Cyronius/Qwen3.6-27B-A2.8B

Ollama
How to use Cyronius/Qwen3.6-27B-A2.8B with Ollama:
```
ollama run hf.co/Cyronius/Qwen3.6-27B-A2.8B
```
Unsloth Desktop

How to use Cyronius/Qwen3.6-27B-A2.8B with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Cyronius/Qwen3.6-27B-A2.8B

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": "Cyronius/Qwen3.6-27B-A2.8B"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use Cyronius/Qwen3.6-27B-A2.8B with Docker Model Runner:
```
docker model run hf.co/Cyronius/Qwen3.6-27B-A2.8B
```

Lemonade

How to use Cyronius/Qwen3.6-27B-A2.8B with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Cyronius/Qwen3.6-27B-A2.8B

Run and chat with the model

lemonade run user.Qwen3.6-27B-A2.8B-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use Cyronius/Qwen3.6-27B-A2.8B with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Cyronius/Qwen3.6-27B-A2.8B

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 Cyronius/Qwen3.6-27B-A2.8B

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use Cyronius/Qwen3.6-27B-A2.8B with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Cyronius/Qwen3.6-27B-A2.8B

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 "Cyronius/Qwen3.6-27B-A2.8B" \
  --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"

Qwen3.6-27B-A2.8B

Pruned, LoRA-healed, and quantized derivative of Qwen/Qwen3.6-35B-A3B, built for fast decode on shared-memory hardware (CPU/iGPU, no discrete GPU required). 10 of 40 language-model layers were removed (1 of every 3 gated-DeltaNet layers), the result was LoRA-healed to recover quality, then quantized to GGUF.

	Qwen3.6-35B-A3B (base)	Qwen3.6-27B-A2.8B (this model)
total params (LM core)	34.7B	26.2B
active params / token	3.45B	2.83B
layers	40	30

Files

Two GGUF builds, both with the multi-token-prediction (MTP) draft head retained for speculative decoding:

file	size	quant	notes
`qwen36-27b-a2.8b-mtp-Q4KM.gguf`	16.49 GB	Q4_K_M, imatrix	balanced — best overall accuracy
`qwen36-27b-a2.8b-mtp-iq3exp-q4head.gguf`	12.01 GB	routed experts → IQ3_S, `lm_head` → Q4_K, rest unchanged	speed-demon — smaller and faster, some accuracy cost

Both share the same imatrix and the same attention / GDN / MTP tensors; the speed-demon build only pushes the routed-expert and output-head tensors to a smaller quant.

Running it

llama.cpp / llama-server, with MTP speculative decoding enabled:

llama-server -m qwen36-27b-a2.8b-mtp-Q4KM.gguf --spec-type draft-mtp --spec-draft-n-max 2

Needs a llama.cpp build with MTP support. Default routed-expert count (k=8) is recommended — dropping it did not give a reproducible speed gain in testing (see Limitations).

Benchmarks

Tool-calling: agentic tool-use cases scored on whether the model calls the right tool with the right arguments, or correctly declines. GSM8K / MMLU: standard accuracy. Perplexity: wikitext-2 test set, 32 chunks, context 512, llama.cpp perplexity tool. Decode: ordinary generation, tokens/sec. Tool decode: generation while producing tool-call output.

model	size	tools	GSM8K	MMLU	ppl	decode	tool decode
balanced (Q4_K_M)	16.49 GB	.95	.80	.733	8.01	36.8	40.7
speed-demon (IQ3_S experts)	12.01 GB	.90	.733	.767	8.29	40.7	44.1
base Q4_K_M (Qwen3.6-35B-A3B, for reference)	21.17 GB	.90	.933	.767	5.50	27.8	27.9

Decode tok/s varied between repeat runs; treat those two columns as directional, not exact. Full result logs are in the GitHub repo's bench/results/.

Limitations

LoRA healing covered roughly half an epoch (18K prompts / 16M tokens). GSM8K is still behind the base model, and perplexity recovered most but not all of the way.
The speed-demon build trades tool accuracy and GSM8K for size and speed.
MTP speculative decoding is not bit-exact vs. plain decoding: a 40-case check scored 37/40 vs. 39/40 for plain decoding. Functionally equivalent, not guaranteed identical.
Lowering routed-expert count (k) below 8 did not give a reproducible speed benefit in testing — not recommended.
Evaluated on a small in-house benchmark (7 tool schemas, GSM8K/MMLU subsets); results may not generalize to other workloads.

How this model was made

Full write-up, code, and result logs: github.com/Cyronius/qwen-prune-heal-pipeline

Short version: 1-in-3 gated-DeltaNet layers were structurally removed from the 40-layer base, then a rank-32 LoRA (~33M trainable params, ~$53 of H200 time) was trained on a mix of general text, tool-use, and GSM8K-style reasoning data to heal the cut, then the result was quantized to GGUF with a calibrated imatrix.

License

Apache-2.0, inherited from the base model, Qwen/Qwen3.6-35B-A3B.

Downloads last month: 95

GGUF

Model size

27B params

Architecture

qwen35moe

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for Cyronius/Qwen3.6-27B-A2.8B

Base model

Qwen/Qwen3.6-35B-A3B

Quantized

(776)

this model