Application images are built on top of runtime images and package a ready-to-run AI application together with the FlagOS stack. Each application is published as a per-backend image, and every image has its own page with the image reference, prerequisites, and launch commands.

Application images are named after a generic vendor rather than the chip brand (for example generic-12.8); base and runtime images keep their real vendor names. The “Built on” reference on each page points at the runtime image the application was built from, under its real name.

Pick an application and a backend to see how to run it:

Megatron training

A ready-to-run environment for training large language models with Megatron-LM. The image already contains FlagOS's build of megatron-core and the full FlagOS software stack — one command on a supported accelerator starts training.

VendorCodeBackendVersion
Ascendascendcann8.5.00.17.1
Ascendascendcann8.5.00.18.2
Ascendascendcann8.5.0-910c0.17.1
Ascendascendcann8.5.0-910c0.18.2
Ascendascendcann9.0.00.17.1
Ascendascendcann9.0.00.18.2
Ascendascendcann9.0.0-910c0.17.1
Ascendascendcann9.0.0-910c0.18.2
Cambriconcambriconneuware4.4.30.17.1
Cambriconcambriconneuware4.7.20.17.1
Cambriconcambriconneuware4.7.20.18.2
Enflameenflametops1.10.60.17.1
Enflameenflametops1.9.100.17.1
Hygonhygondtk26.040.17.1
Hygonhygondtk26.040.18.2
Iluvatariluvatarcorex4.4.00.17.1
Iluvatariluvatarcorex4.5.00.17.1
Kunlunxinkunlunxinxre5.37.10.17.1
MetaXmetaxmaca3.7.2.10.17.1
MetaXmetaxmaca3.8.1.30.17.1
MetaXmetaxmaca3.8.1.30.18.2
Moore Threadsmthreadsmusa4.3.60.17.1
Moore Threadsmthreadsmusa5.2.00.17.1
Genericgeneric12.80.17.1
Genericgeneric12.80.18.2
Genericgeneric13.30.17.1
Genericgeneric13.30.18.2
Sunrisesunrisetangrt1.2.00.17.1
TsingMicrotsingmicrotsm2606100.17.1

Megatron RL

A ready-to-run environment for reinforcement learning with Megatron-LM. The image already contains FlagOS's build of megatron-core with reinforcement-learning support, so a GRPO-style training run needs no environment setup of its own.

VendorCodeBackendVersion
Ascendascendcann8.5.00.17.1
Ascendascendcann8.5.0-910c0.17.1
Ascendascendcann9.0.00.17.1
Ascendascendcann9.0.0-910c0.17.1
Cambriconcambriconneuware4.4.30.17.1
Cambriconcambriconneuware4.7.20.17.1
Cambriconcambriconneuware4.7.20.18.2
Enflameenflametops1.10.60.17.1
Enflameenflametops1.9.100.17.1
Hygonhygondtk26.040.17.1
Hygonhygondtk26.040.18.2
Iluvatariluvatarcorex4.4.00.17.1
Iluvatariluvatarcorex4.5.00.17.1
Kunlunxinkunlunxinxre5.37.10.17.1
MetaXmetaxmaca3.7.2.10.17.1
MetaXmetaxmaca3.8.1.30.17.1
MetaXmetaxmaca3.8.1.30.18.2
Moore Threadsmthreadsmusa4.3.60.17.1
Moore Threadsmthreadsmusa5.2.00.17.1
Genericgeneric12.80.17.1
Genericgeneric12.80.18.2
Genericgeneric13.30.17.1
Genericgeneric13.30.18.2
Sunrisesunrisetangrt1.2.00.17.1
TsingMicrotsingmicrotsm2606100.17.1

vLLM

A ready-to-run inference service for large language models. The image already contains vLLM, its FlagOS plugin and the full FlagOS software stack — start it on a supported accelerator and it serves an OpenAI-compatible API.

VendorCodeBackendVersion
Ascendascendcann8.5.00.20.2
Ascendascendcann8.5.00.24.0
Ascendascendcann8.5.0-910c0.20.2
Ascendascendcann8.5.0-910c0.24.0
Ascendascendcann9.0.00.20.2
Ascendascendcann9.0.00.24.0
Ascendascendcann9.0.0-910c0.20.2
Ascendascendcann9.0.0-910c0.24.0
Cambriconcambriconneuware4.4.30.20.2
Cambriconcambriconneuware4.4.30.24.0
Cambriconcambriconneuware4.7.20.20.2
Cambriconcambriconneuware4.7.20.24.0
Enflameenflametops1.10.60.20.2
Enflameenflametops1.10.60.24.0
Enflameenflametops1.9.100.20.2
Enflameenflametops1.9.100.24.0
Hygonhygondtk26.040.20.2
Hygonhygondtk26.040.24.0
Iluvatariluvatarcorex4.4.00.20.2
Iluvatariluvatarcorex4.4.00.24.0
Iluvatariluvatarcorex4.5.00.20.2
Iluvatariluvatarcorex4.5.00.24.0
Kunlunxinkunlunxinxre5.37.10.20.2
Kunlunxinkunlunxinxre5.37.10.24.0
MetaXmetaxmaca3.7.2.10.20.2
MetaXmetaxmaca3.7.2.10.24.0
MetaXmetaxmaca3.8.1.30.20.2
MetaXmetaxmaca3.8.1.30.24.0
Moore Threadsmthreadsmusa4.3.60.20.2
Moore Threadsmthreadsmusa4.3.60.24.0
Moore Threadsmthreadsmusa5.2.00.20.2
Moore Threadsmthreadsmusa5.2.00.24.0
Genericgeneric12.80.20.2
Genericgeneric12.80.24.0
Genericgeneric13.30.20.2
Genericgeneric13.30.24.0
Sunrisesunrisetangrt1.2.00.20.2
Sunrisesunrisetangrt1.2.00.24.0
TsingMicrotsingmicrotsm2606100.20.2
TsingMicrotsingmicrotsm2606100.24.0

SGLang

A ready-to-run inference service for large language models, with structured output built in. The image already contains SGLang, its FlagOS plugin and the full FlagOS software stack — start it on a supported accelerator and it serves an API.

VendorCodeBackendVersion
Ascendascendcann8.5.00.5.18
Ascendascendcann9.0.00.5.18
Cambriconcambriconneuware4.4.30.5.18
Cambriconcambriconneuware4.7.20.5.18
Enflameenflametops1.10.60.5.18
Enflameenflametops1.9.100.5.18
Hygonhygondtk26.040.5.18
Iluvatariluvatarcorex4.4.00.5.18
Iluvatariluvatarcorex4.5.00.5.18
Kunlunxinkunlunxinxre5.37.10.5.18
MetaXmetaxmaca3.7.2.10.5.18
MetaXmetaxmaca3.8.1.30.5.18
Moore Threadsmthreadsmusa4.3.60.5.18
Moore Threadsmthreadsmusa5.2.00.5.18
Sunrisesunrisetangrt1.2.00.5.18