Skip to content

Hardware support

Use this page to check whether a prebuilt binary exists for your accelerator. If you are installing on a normal workstation or server, start with the quickstart and let the installer choose.

PlatformAccelerationPrebuilt binary
Linux x86_64 + NVIDIA GPUCUDA (Ampere and newer)yes, per compute capability and driver CUDA support
Linux aarch64 + NVIDIA GPUCUDA (Grace: GH200/GB200/GB10)yes, sm90/100/121, per driver CUDA support
Apple Silicon (macOS arm64)Metalyes
Linux x86_64 / aarch64, no GPUCPUyes
Windows x86_64CPUyes
Intel Mac, unlisted GPUsource buildno

The minimum supported NVIDIA GPU is Ampere (compute capability 8.0). Turing (sm75: RTX 20-series, GTX 16-series, Tesla T4) and older are not supported by current prebuilts.

Compute capabilityArchitectureRepresentative GPUs
8.0Ampere (datacenter)A100, A30
8.6Ampere (consumer)RTX 3090/3080/3070/3060, A40, A10
8.9AdaRTX 4090/4080, L40, L4
9.0HopperH100, H200
10.0Blackwell (datacenter)B200, GB200
12.0Blackwell (consumer)RTX 5090/5080
12.1GB10DGX Spark

CUDA artifact names encode both the toolkit lane and compute capability:

mistralrs-cuda128-sm90-aarch64-unknown-linux-gnu.tar.gz

The installer chooses the newest published lane that the installed NVIDIA driver can load.

Driver reportsArtifact laneNotes
CUDA 13.3+ on Hopper / sm90cuda133cuTile lane for Hopper
CUDA 13.2+ on Ampere/Ada / sm80, sm86, sm89cuda132cuTile lane for Ampere and Ada
CUDA 13.1+ on Blackwell / sm100, sm120, sm121cuda131cuTile lane for Blackwell and GB10
CUDA 13.0+cuda130CUDA 13 baseline lane
CUDA 12.9+ on GB10 / sm121cuda129needed because CUDA 12.8 does not target sm121
CUDA 12.8+cuda128baseline lane for Ampere and newer
ArchitectureCompute capabilities
x86_6480, 86, 89, 90, 100, 120
aarch64 (NVIDIA Grace: GH200, GB200, GB10/DGX Spark)90, 100, 121
Asset tokenBuilt withMinimum nvidia-smi CUDA versionPublished compute capabilitiescuTile
cuda128CUDA 12.8.112.8x86_64: 80, 86, 89, 90, 100, 120; aarch64: 90, 100no
cuda129CUDA 12.9.112.9aarch64: 121no
cuda130CUDA 13.0.013.0x86_64: 80, 86, 89, 90, 100, 120; aarch64: 90, 100, 121no
cuda131CUDA 13.1.213.1x86_64: 80, 86, 89, 90, 100, 120; aarch64: 90, 100, 121sm100, sm120, sm121
cuda132CUDA 13.2.013.2x86_64: 80, 86, 89sm80, sm86, sm89
cuda133CUDA 13.3.013.3x86_64: 90; aarch64: 90sm90

Each artifact bundles the CUDA runtime libraries it needs, so no CUDA toolkit is required at runtime. The installed NVIDIA driver still has to be new enough for that artifact’s toolkit lane.

The same compatibility lanes are used by the Docker images and CUDA Python wheels. See Python getting started for wheel install commands.

FeatureRequirement
flash-attn (v2)compute capability 8.0+
flash-attn-v3Hopper (9.0)
FP8 matmulcompute capability 8.9+
cuTile MoE backendAmpere/Ada (8.x) with CUDA >= 13.2; Hopper (9.0) with CUDA >= 13.3; Blackwell+ (10.x/12.x) with CUDA >= 13.1
CUTLASS MoE backendcompute capability 8.0+

See cargo features for the feature flags and MoE expert backends for backend selection.