Description
Mage-Flow-Turbo-Native-Inference is a portable native inference and deployment stack for the Mage-Flow-Turbo image diffusion model. The repository does not train or modify model weights: its Python layer handles configuration, model and runtime identity verification, CLI and REST orchestration, lifecycle control, telemetry and evidence collection, while the actual image generation runs through the native stable-diffusion.cpp sd-cli runtime rather than a PyTorch or Transformers inference loop.
Read More
Version 1.0.0 ships two model profiles through that same native runtime. The canonical q8-reference profile pairs GGUF Q8_0 diffusion weights with a Qwen3-VL-4B GGUF text encoder and a SafeTensors VAE. The bf16-safetensors profile uses BF16 SafeTensors diffusion weights with the same text encoder and VAE. Both profiles run on CPU as well as on CUDA (the cuda0 device), and every model component's identity is checked against its SHA-256 hash, along with a pinned native runtime commit, before real inference happens.
A mageflow-native CLI verifies, builds and runs each profile, and a local REST API, which binds to 127.0.0.1 by default, exposes health and readiness checks, a service-info endpoint, an image-generation endpoint, and an endpoint for retrieving generated PNG artifacts.
The project publishes a strict 2x2 benchmark protocol comparing both model profiles on CPU and on Kaggle T4/T4x2 GPUs, using a fixed prompt, seed, step count and CFG value, and a resolution matrix from 512 up to 1024 pixels. Results are recorded as checksum-protected evidence tied to the exact source revision, model identities and runtime commit, rather than pasted as plain numbers, and a public Kaggle notebook plus documented Kaggle Dataset and Model identities let others reproduce the same runs on either accelerator.
The project is built for developers who want a reproducible, benchmarked way to run the Mage-Flow-Turbo model on CPU or CUDA without standing up a full Python ML stack. It is released under the MIT license and is free to use. It has no anime-specific content, presets or tooling; it is a general native inference stack for the underlying model.
- CategoryAI Image Generators
- Pricing ModelOpen source
- DeploymentSelf-hosted
- PlatformsWeb
Mage-Flow-Turbo-Native-Inference's Core Features
Native stable-diffusion.cpp sd-cli execution instead of a PyTorch/Transformers inference loop
Two supported model profiles: Q8 GGUF diffusion (canonical/default) and BF16 SafeTensors diffusion
CPU and CUDA (cuda0) backends for both model profiles
SHA-256 identity verification of model weights and the native runtime before inference
mageflow-native CLI for verifying, building and running the runtime
Local REST API with health checks and a POST endpoint for image generation
Published strict 2x2 CPU vs. Kaggle T4/T4x2 benchmark protocol with a fixed prompt and resolution matrix
Public Kaggle notebook and documented Dataset/Model identities for reproducing benchmark runs
Structured evidence recording of source revision, model identities, runtime commit and generation metadata
Getting Started with Mage-Flow-Turbo-Native-Inference
Install locally: Run pip install -e . in a Python environment on Linux.
Build the runtime: Run mageflow-native runtime build with the cpu or cuda backend.
Verify the profile: Run mageflow-native doctor with your chosen manifest to check model and runtime identities.
Generate an image: Call mageflow-native verify, or send a request to the local /v1/images/generate REST endpoint.
Reproduce benchmarks: Open the Kaggle notebook and attach the matching runtime Dataset and model Models for your accelerator.
Mage-Flow-Turbo-Native-Inference's Use Cases
- Local image generation
- CPU vs. GPU benchmarking
- Self-hosted image API
