Instructions to use marlalabsAI/Qwen3.8-27B-SX8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use marlalabsAI/Qwen3.8-27B-SX8 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 marlalabsAI/Qwen3.8-27B-SX8
# Run inference directly in the terminal:
llama cli -hf marlalabsAI/Qwen3.8-27B-SX8

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf marlalabsAI/Qwen3.8-27B-SX8
# Run inference directly in the terminal:
llama cli -hf marlalabsAI/Qwen3.8-27B-SX8

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 marlalabsAI/Qwen3.8-27B-SX8
# Run inference directly in the terminal:
./llama-cli -hf marlalabsAI/Qwen3.8-27B-SX8

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 marlalabsAI/Qwen3.8-27B-SX8
# Run inference directly in the terminal:
./build/bin/llama-cli -hf marlalabsAI/Qwen3.8-27B-SX8

Use Docker

docker model run hf.co/marlalabsAI/Qwen3.8-27B-SX8

LM Studio
Jan

vLLM

How to use marlalabsAI/Qwen3.8-27B-SX8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "marlalabsAI/Qwen3.8-27B-SX8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "marlalabsAI/Qwen3.8-27B-SX8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/marlalabsAI/Qwen3.8-27B-SX8

Ollama
How to use marlalabsAI/Qwen3.8-27B-SX8 with Ollama:
```
ollama run hf.co/marlalabsAI/Qwen3.8-27B-SX8
```
Unsloth Desktop

How to use marlalabsAI/Qwen3.8-27B-SX8 with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf marlalabsAI/Qwen3.8-27B-SX8

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": "marlalabsAI/Qwen3.8-27B-SX8"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use marlalabsAI/Qwen3.8-27B-SX8 with Docker Model Runner:
```
docker model run hf.co/marlalabsAI/Qwen3.8-27B-SX8
```

Lemonade

How to use marlalabsAI/Qwen3.8-27B-SX8 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull marlalabsAI/Qwen3.8-27B-SX8

Run and chat with the model

lemonade run user.Qwen3.8-27B-SX8-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use marlalabsAI/Qwen3.8-27B-SX8 with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf marlalabsAI/Qwen3.8-27B-SX8

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 marlalabsAI/Qwen3.8-27B-SX8

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use marlalabsAI/Qwen3.8-27B-SX8 with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf marlalabsAI/Qwen3.8-27B-SX8

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 "marlalabsAI/Qwen3.8-27B-SX8" \
  --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"

Qwen3.8-27B-SX8 — S-X8 v4.3 quantized variant of Qwen3.8-27B

S-X8 v4.3 is a weight quantization format for large language models: 30 bytes per block = 7.50 bpp (fully accounted), FP16-level quality, and a portable byte-aligned decoder. This repository contains the Qwen3.8-27B quantized to S-X8 v4.3 — a larger variant of the same quantization applied to Qwen3.5-4B.

Format validated on Qwen3.5-4B

The S-X8 v4.3 format is validated on Qwen3.5-4B (single RTX 5060 Ti). See the 4B repository for the full benchmark table and the paper:

Metric	FP16	S-X8 v4.3 (4B)	Q8_0
PPL wikitext-2 (PCA runtime)	10.2090	10.2267 (+0.17%)	10.4540 (+2.40%)
Winogrande_s	0.5746	0.5722	0.5746
Decode, real-world (llama.cpp fork)	—	63.79 tok/s	40–54 tok/s
VRAM decode M=1	—	3.720 GB	4.48 GB

Paper (Zenodo, DOI): 10.5281/zenodo.21922640

This 27B model

Quantized with the same pipeline as the 4B: per-tensor S-X8 v4.3, 666 tensors, CosSim 0.999753 (mean over all tensors), 7.50 bpp, 26.1 GB container.

Validated functionally: the model runs and answers correctly in a local test (same protocol as the 4B model). Formal 27B benchmarks (PPL, Winogrande, ARC, MMLU) are pending — to be measured on a larger GPU (we will update when measured). The larger base model is expected to deliver better absolute quality (same format, more capable base).

Two containers, one model

The S-X8 model exists in two containers — same weights inside:

File	Container	Used by
`Qwen3.8-27B-SX8v43.sx8`	native S-X8 v4.3 format (complete: config + quantized 2D + 1D + MTP + vision)	our Python runtime (`scripts/infer_sx8.py`, CUDA kernels)
`Qwen3.8-27B-SX8v43.gguf`	GGUF (the container llama.cpp reads) with S-X8 weights inside — native type `GGML_TYPE_SX8 = 41`	llama.cpp fork (`run_llama_chat.sh`)

Think of it like a .zip: the container is the same, what's inside can be anything. The GGUF is the same S-X8 inside the container llama.cpp reads. MTP lives in the .sx8v43 (excluded from the GGUF, same criterion as the 4B).

Quick start

Vía A — llama.cpp fork (fast chat):

./run_llama_chat.sh build   # one time (clones llama.cpp, applies the patch, compiles)
./run_llama_chat.sh         # chat

GPU ≥ 32 GB: NGL=999 ./run_llama_chat.sh (everything in VRAM)
GPU 16 GB: NGL=28 ./run_llama_chat.sh (28 layers GPU + rest CPU), or unified memory: GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 NGL=999 ./run_llama_chat.sh (100% GPU compute, PCIe spill)

Vía B — Python runtime with our CUDA kernels (native container):

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python3 scripts/infer_sx8.py --container Qwen3.8-27B-SX8v43.sx8

Full case-by-case guide: QUICK-DEPLOY.md (GPUs 8–32+ GB, CPU-only, Windows/WSL2).

Repository layout

├── Qwen3.8-27B-SX8v43.sx8      native S-X8 container (complete model)
├── Qwen3.8-27B-SX8v43.gguf      S-X8 inside GGUF (llama.cpp)
├── tokenizer/                  (tokenizer.json, tokenizer_config.json, vocab.json, merges.txt, chat_template.jinja)
├── scripts/                    quantization (27B) · containers · runtime (eval_common, infer_sx8) · kernels
├── cuda/                       S-X8 CUDA kernels (decode1_v3, fused_wmma, gemm_compact, embed_v44...)
├── llama-cpp-sx8.patch         llama.cpp fork: S-X8 as native type (MMVQ decode + MMQ prompt + AVX2 CPU)
├── QUICK-DEPLOY.md             deployment guide by GPU case
├── SX8_FLASH_V4_3_SPEC.md      block format spec (30 B/block)
├── SX8_FLASH_V4_3_CONTAINER.md container format spec (v1.1 SXT1)
├── S-X-METHODOLOGY.md          methodology
└── CHANGELOG.md

Transparency

The conceptual seeds of S-X8 come from an independent mathematical analysis of the image of the Shroud of Turin; the full study is published in the project repo with its own re-verification: github.com/MarlaLabsAI/sx8-quantization → shroud-turin-study/. See the paper's Appendix A and IDEA-PROVENANCE.md. The format itself is validated empirically in this repository; the study is disclosed only as the source of inspiration, with full honesty about what can and cannot be corroborated.

License and attribution

Apache-2.0. The S-X8 format, kernels and scripts in this repository are original work (Copyright (C) 2026 Martí Vidal Leandro), not a Qwen work. The quantized model uses Qwen3.8-27B by the Qwen Team (Alibaba Group), Apache-2.0 (model card) as the base model; only the weights were quantized to the S-X8 v4.3 format (see NOTICE), no other modifications.

How to cite

If you use this work, please cite:

Vidal Leandro, M. (2026). S-X8 v4.3: A 7.50-Bits-Per-Weight Quantization Format with FP16-Level Quality and Portable Decoding. Zenodo. https://doi.org/10.5281/zenodo.21922640

Contact

Looking for collaborations, integration help or opportunities? Reach out:

LinkedIn (profile): https://www.linkedin.com/in/vidalmarti/
LinkedIn (company): https://www.linkedin.com/company/marlalabs/
X / Twitter: https://x.com/MarlaLabsAI
Email: info@marlalabs.com
Web: https://marlalabs.com

Support

If this work helps you, consider a donation: Ko-fi · GitHub Sponsors — details in FUNDING.yml.

Deploy with an AI agent

Want another AI (Claude Code, opencode, aider...) to handle the whole deployment for you — on any OS and hardware, within physical limits? Copy the prompt from DEPLOY-AI.md and give it to the agent: it will detect your system (OS, GPU, RAM), install what's needed, download, run and verify the model on its own.

License / NOTICE

Apache-2.0. The S-X8 format is original work of the author; this model is a quantized variant of Qwen3.8-27B (Qwen Team, Alibaba Group, Apache-2.0) — only the weights were quantized to S-X8 v4.3, no other modifications. See NOTICE.

Qwen3.8-27B — S-X8 v4.3 (7,50 bpp) [ES]

S-X8 v4.3 es un formato de cuantización de pesos para modelos de lenguaje: 30 bytes por bloque = 7,50 bpp (completamente contabilizado), calidad a nivel FP16, y un decodificador portable alineado a byte. Este repositorio contiene el Qwen3.8-27B cuantizado a S-X8 v4.3 — una variante más grande de la misma cuantización aplicada a Qwen3.5-4B.

Formato validado en Qwen3.5-4B

El formato S-X8 v4.3 está validado en Qwen3.5-4B (una RTX 5060 Ti). Consulta el repositorio del 4B para la tabla completa de benchmarks y el paper:

Métrica	FP16	S-X8 v4.3 (4B)	Q8_0
PPL wikitext-2 (runtime PCA)	10,2090	10,2267 (+0,17%)	10,4540 (+2,40%)
Winogrande_s	0,5746	0,5722	0,5746
Decode real (fork llama.cpp)	—	63,79 tok/s	40–54 tok/s
VRAM decode M=1	—	3,720 GB	4,48 GB

Paper (Zenodo, DOI): 10.5281/zenodo.21922640

Este modelo 27B

Cuantizado con el mismo pipeline que el 4B: S-X8 v4.3 por tensor, 666 tensores, CosSim 0,999753 (media sobre todos los tensores), 7,50 bpp, 26,1 GB de contenedor.

Validado funcionalmente: el modelo se ejecuta y responde correctamente en una prueba local (mismo protocolo que el modelo 4B). Los benchmarks formales del 27B (PPL, Winogrande, ARC, MMLU) están pendientes — se medirán en una GPU más grande (actualizaremos cuando se midan). Se espera que el modelo base más grande ofrezca mejor calidad absoluta (mismo formato, base más capaz).

Dos envases, un modelo

El modelo S-X8 existe en dos envases — los mismos pesos dentro:

Archivo	Envase	Usado por
`Qwen3.8-27B-SX8v43.sx8`	formato nativo S-X8 v4.3 (completo: config + 2D cuantizado + 1D + MTP + visión)	nuestro runtime Python (`scripts/infer_sx8.py`, kernels CUDA)
`Qwen3.8-27B-SX8v43.gguf`	GGUF (el envase que lee llama.cpp) con pesos S-X8 dentro — tipo nativo `GGML_TYPE_SX8 = 41`	fork llama.cpp (`run_llama_chat.sh`)

Piénsalo como un .zip: el envase es el mismo, lo de dentro puede ser lo que sea. El GGUF es el mismo S-X8 dentro del envase que lee llama.cpp. El MTP vive en el .sx8v43 (excluido del GGUF, mismo criterio que el 4B).

Inicio rápido

Vía A — fork llama.cpp (chat rápido):

./run_llama_chat.sh build   # una vez (clona llama.cpp, aplica el patch, compila)
./run_llama_chat.sh         # chat

GPU ≥ 32 GB: NGL=999 ./run_llama_chat.sh (todo en VRAM)
GPU 16 GB: NGL=28 ./run_llama_chat.sh (28 capas GPU + resto CPU), o memoria unificada: GGML_CUDA_ENABLE_UNIFIED_MEMORY=1 NGL=999 ./run_llama_chat.sh (cómputo 100% GPU, spill por PCIe)

Vía B — runtime Python con nuestros kernels CUDA (envase nativo):

python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
python3 scripts/infer_sx8.py --container Qwen3.8-27B-SX8v43.sx8

Guía completa caso por caso: QUICK-DEPLOY.md (GPUs 8–32+ GB, solo CPU, Windows/WSL2).

Estructura del repositorio

├── Qwen3.8-27B-SX8v43.sx8      contenedor nativo S-X8 (modelo completo)
├── Qwen3.8-27B-SX8v43.gguf      S-X8 dentro de GGUF (llama.cpp)
├── tokenizer/                  (tokenizer.json, tokenizer_config.json, vocab.json, merges.txt, chat_template.jinja)
├── scripts/                    cuantización (27B) · contenedores · runtime (eval_common, infer_sx8) · kernels
├── cuda/                       kernels CUDA S-X8 (decode1_v3, fused_wmma, gemm_compact, embed_v44...)
├── llama-cpp-sx8.patch         fork llama.cpp: S-X8 como tipo nativo (decode MMVQ + prompt MMQ + AVX2 CPU)
├── QUICK-DEPLOY.md             guía de despliegue por caso de GPU
├── DEPLOY-AI.md                prompt para que otra IA despliegue todo de forma autónoma
├── SX8_FLASH_V4_3_SPEC.md      especificación del bloque (30 B/bloque)
├── SX8_FLASH_V4_3_CONTAINER.md especificación del contenedor (v1.1 SXT1)
├── S-X-METHODOLOGY.md          metodología
└── CHANGELOG.md

Transparencia

Las semillas conceptuales de S-X8 provienen de un análisis matemático independiente de la imagen de la Sábana Santa de Turín; el estudio completo se publica en el repo del proyecto con su propia re-verificación: github.com/MarlaLabsAI/sx8-quantization → shroud-turin-study/. Ver el Apéndice A del paper y IDEA-PROVENANCE.md. El formato en sí está validado empíricamente en este repositorio; el estudio se divulga solo como fuente de inspiración, con total honestidad sobre lo que se puede y no se puede corroborar.

Licencia y atribución

Apache-2.0. El formato S-X8, los kernels y los scripts de este repositorio son obra original (Copyright (C) 2026 Martí Vidal Leandro), no un trabajo de Qwen. El modelo cuantizado usa Qwen3.8-27B de Qwen Team (Alibaba Group), Apache-2.0 (model card) como modelo base; solo se cuantizaron los pesos al formato S-X8 v4.3 (ver NOTICE), sin otras modificaciones.

Cómo citar

Si usas este trabajo, cítalo como:

Vidal Leandro, M. (2026). S-X8 v4.3: Un formato de cuantización a 7,50 bits por peso con calidad de FP16 y decodificación portable. Zenodo. https://doi.org/10.5281/zenodo.21922640

Contacto

¿Buscas colaboración, ayuda con la integración u oportunidades? Contacta:

LinkedIn (perfil): https://www.linkedin.com/in/vidalmarti/
LinkedIn (empresa): https://www.linkedin.com/company/marlalabs/
X / Twitter: https://x.com/MarlaLabsAI
Email: info@marlalabs.com
Web: https://marlalabs.com

Apoyo

Si este trabajo te ayuda, considera una donación: Ko-fi · GitHub Sponsors — detalles en FUNDING.yml.

Despliegue con un agente de IA

¿Quieres que otra IA (Claude Code, opencode, aider...) se encargue de todo el despliegue por ti — en cualquier sistema operativo y hardware, dentro de los límites físicos? Copia el prompt de DEPLOY-AI.md y entrégaselo al agente: detectará tu sistema (SO, GPU, RAM), instalará lo necesario, descargará, ejecutará y verificará el modelo por su cuenta.

Archivos de uso en este repo (autosuficiente)

llama-cpp-sx8.patch — patch del fork de llama.cpp (aplicar sobre el commit 7c203670f)
SX8_FLASH_V4_3_SPEC.md · SX8_FLASH_V4_3_CONTAINER.md — spec del formato y del contenedor
S-X-METHODOLOGY.md — autoría de la metodología
IDEA-PROVENANCE.md — procedencia de la idea (estudio Sábana Santa)
DEPLOY-AI.md · QUICK-DEPLOY.md — guías de despliegue
scripts/ · cuda/ — runtime Python y kernels CUDA S-X8
README.md (EN) · README-ES.md (ES) — documentación

Downloads last month: 185

GGUF

Model size

27B params

Architecture

qwen35

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for marlalabsAI/Qwen3.8-27B-SX8

Base model

Qwen/Qwen3.8-27B

Quantized

(876)

this model