Instructions to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M",
	filename="qwen35-0.8b-spectralquant-calib360-q4_k_m.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M 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 Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

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 Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

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 Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

Use Docker

docker model run hf.co/Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

LM Studio
Jan

vLLM

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

Ollama
How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with Ollama:
```
ollama run hf.co/Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M
```

Unsloth Studio

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M to start chatting

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M: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 Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

Run Hermes

hermes

Atomic Chat new
Docker Model Runner
How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with Docker Model Runner:
```
docker model run hf.co/Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M
```

Lemonade

How to use Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.5-0.8B-SpectralQuant-Q4_K_M-Q4_K_M

List all available models

lemonade list

Qwen3.5 0.8B SpectralQuant Q4_K_M

SpectralQuant Q4_K_M is a compact GGUF release of Qwen/Qwen3.5-0.8B, built with a new calibration-aware quantization approach. Instead of treating Q4 compression as simple local rounding, SpectralQuant shapes the quantized representation around behaviorally important directions, keeping the normal Q4_K_M footprint while preserving substantially more of the BF16 reference behavior.

A detailed technical blog post describing the method and research path is planned soon.

Highlights

4.52 BPW fixed-footprint GGUF: 435,896,640 bytes / 415.7 MiB.
96.5% heldout120 BF16-gap recovery versus llama.cpp pure Q4_K_M.
Lower prompt loss than tested Unsloth Q4_K_S, Q4_K_M, IQ4_NL, and IQ4_XS while using fewer bytes.
C4 validation improves over llama.cpp pure Q4_K_M at the same Q4_K_M footprint.
No FP-kept modules, no mixed-precision sidecar, and no larger dynamic quant format.

Model Details

Item	Value
Base model	`Qwen/Qwen3.5-0.8B`
Format	GGUF
Quantization	SpectralQuant Q4_K_M
File	`qwen35-0.8b-spectralquant-calib360-q4_k_m.gguf`
Size	`435,896,640` bytes
SHA256	`ae3c3e6dbb3d08c83d12e12c2f67bf63527f39e5090ce5c0eb12eacd5417f352`
License	Apache-2.0

This release is quantized from Qwen/Qwen3.5-0.8B. It is not quantized directly from Qwen/Qwen3.5-0.8B-Base; any Base -> Qwen3.5-0.8B -> Quantized lineage shown by the Hub reflects the upstream Qwen model relationship.

Method Overview

SpectralQuant is a calibration-aware Q4_K_M quantization approach. At a high level, it shapes quantization error around behaviorally important directions instead of treating every local rounding error equally. The goal is simple: keep the familiar Q4_K_M deployment footprint while retaining more of the model behavior users normally associate with larger quantizations.

A detailed technical blog post describing the method and research path is planned soon.

Evaluation

Lower loss is better. BPW is estimated from file size relative to llama.cpp pure Q4_K_M at 4.52 BPW. BF16 is included as a full-precision reference.

Model	BPW est.	Size MiB	convergence60 ?	heldout120 ?
BF16 reference	16.01	1446.5	2.2682	2.9809
Unsloth UD-Q4_K_XL	5.79	532.9	2.2833	2.9913
SpectralQuant Q4_K_M	4.52	415.7	2.2509	2.9961
Unsloth IQ4_NL	5.26	483.4	2.3289	3.0484
Unsloth Q4_K_M	5.52	507.8	2.3268	3.0510
Unsloth Q4_K_S	5.27	484.6	2.3126	3.0700
Unsloth IQ4_XS	5.11	469.8	2.3869	3.1061
llama.cpp pure Q4_K_M	4.52	415.7	2.7404	3.4135

BF16 Gap Recovery

Suite	Pure Q4_K_M loss	BF16 reference loss	SpectralQuant loss	Recovery vs BF16 gap
convergence60	2.740441	2.268226	2.250946	100.00%
heldout120	3.413494	2.980932	2.996070	96.50%

C4 Validation

Suite	llama.cpp pure Q4_K_M	SpectralQuant Q4_K_M	Unsloth Q4_K_M
C4 validation 64x256	3.3014	3.2874	3.2574

Quickstart

llama-cli -m qwen35-0.8b-spectralquant-calib360-q4_k_m.gguf   -p "Explain quantization in two sentences."   -n 80

llama-server -m qwen35-0.8b-spectralquant-calib360-q4_k_m.gguf -c 4096

Notes

Release metrics are prompt-loss and C4 validation results from fixed evaluation suites.
The main claim is bounded to this release table and same-footprint Q4_K_M behavior.
Larger or dynamic quantizations can still win in some settings; evaluate on your workload before deployment.
The base model and referenced GGUF source list Apache-2.0 licensing.

Attribution

Base model: Qwen/Qwen3.5-0.8B
Reference GGUF source: unsloth/Qwen3.5-0.8B-GGUF
Quantization: SpectralQuant Q4_K_M

Downloads last month: -

GGUF

Model size

0.8B params

Architecture

qwen35

Hardware compatibility

4-bit

Model tree for Spectral-Labs25/Qwen3.5-0.8B-SpectralQuant-Q4_K_M

Base model

Qwen/Qwen3.5-0.8B-Base

Finetuned

Qwen/Qwen3.5-0.8B

Quantized

(157)

this model