---

title: 'Spice Runtime Distributions' sidebar_label: 'Distributions' sidebar_position: 31 description: 'Distribution variants of the Spice runtime for different use cases and deployment scenarios, including data-only, GPU-accelerated, NAS, and allocator variants.' pagination_prev: null pagination_next: null

The Spice open source project provides multiple distribution variants to support different use cases and deployment scenarios.

:::note The Spice runtime is 64-bit only. 32-bit platforms are not supported. :::

Image Channels

Channel	Image	Description
DockerHub (opens in a new tab)	spiceai/spiceai	Official release images
GitHub Container Registry (opens in a new tab)	ghcr.io/spiceai/spiceai	Official release images
GitHub Container Registry (Nightly) (opens in a new tab)	ghcr.io/spiceai/spiceai-nightly	Nightly builds with additional variants
AWS Marketplace (opens in a new tab)	—	Enterprise image
Azure Marketplace	—	Enterprise image (coming soon)
Spice Cloud Platform (opens in a new tab)	—	Uses Enterprise image
Spice.ai Enterprise (opens in a new tab)	—	Uses Enterprise image

:::note Some variant distributions are only available in nightly images (data) or exclusively through the Spice Cloud Platform and Spice.ai Enterprise (opens in a new tab) (NAS, CUDA, allocator variants). :::

Supported Platforms and Hardware Requirements

Platform	Architecture	Minimum CPU Features	Build Prerequisites
Linux	x86_64	AVX2, FMA, BMI1/2, LZCNT, POPCNT	—
Linux	aarch64 (arm64)	NEON, FP16 (FEAT_FP16), FHM (FEAT_FHM)	clang, lld
macOS	aarch64 (Apple Silicon)	Native (build host)	—
Windows	x86_64 (MSVC)	—	MSVC toolchain

:::note Windows support is CLI (spice) only. The runtime daemon (spiced) is not supported on Windows natively — use WSL instead. :::

Distribution Availability

Distribution / Variant	Image Tag	Open Source	Spice Cloud	Enterprise
Default (Data + AI)	latest	✅	✅	✅
Data-only	latest-data	Nightly only	✅	✅
NFS connector	—	Local build only	❌	✅
Metal (macOS)	—	Local build only	✅	✅
CUDA (Linux)	latest-cuda	Local build only	✅	✅
Allocator variants	latest-{jemalloc,mimalloc,sysalloc}	Local build only	✅	✅
ODBC connector	—	Local build only	✅	✅
ScyllaDB connector	—	Local build only	✅	✅
Elasticsearch connector	—	Local build only	✅	✅

Default Distribution

The default distribution includes the standard spiced feature set, including AI/ML model support. This is the recommended distribution for most users.

Included Features:

• All standard data connectors (PostgreSQL, MySQL, DuckDB, SQLite, ClickHouse, etc.)
• Embedded data accelerators (Spice Cayenne, DuckDB, SQLite)
• AI/ML model inference (LLMs, embeddings)
• Search capabilities (Vector and BM-25 Full-Text-Search)
• Default memory allocator (snmalloc)

Not included by default:

• The ODBC (odbc), Elasticsearch (elasticsearch), NFS (nfs), and ScyllaDB (scylladb) data connectors — see Additional Connectors (opens in a new tab)

:::note The PostgreSQL data accelerator is only available in nightly builds. The PostgreSQL data connector is included in all distributions. :::

Installation:

Docker:

Data Distribution

The data distribution excludes AI/ML model support, resulting in a smaller binary size and reduced attack surface. Use this when data federation and acceleration capabilities are needed without AI features.

:::note Open Source: Available in nightly builds only. Cloud Platform & Enterprise (opens in a new tab): Production-ready data distribution available. :::

Included Features:

• The default distribution's data connectors, minus ADBC
• All data accelerators
• Default memory allocator (snmalloc)

Excluded Features:

• AI/ML model inference
• LLM support
• Embedding models

Docker (Nightly):

Local Build:

:::note make install-data-only builds from SPICED_DATA_FEATURES in the Makefile — a hand-maintained list that is not the spiced default feature set with models removed, and that differs from it in both directions. Compared with the default distribution it omits the ADBC (opens in a new tab) data connector and the AWS Secrets Manager (opens in a new tab), Azure Key Vault (opens in a new tab) and keyring (opens in a new tab) secret stores, and it adds the PostgreSQL data accelerator (opens in a new tab) and acceleration snapshots (opens in a new tab). The environment (opens in a new tab) and Kubernetes (opens in a new tab) secret stores are not feature-gated and are available in every build.

A feature added to the default set does not reach make install-data-only until it is added to SPICED_DATA_FEATURES too, so recompute the difference from that list and the default = [...] array in bin/spiced/Cargo.toml rather than assuming the two track each other. :::

GPU-Accelerated Distributions

Metal (macOS)

For macOS systems with Apple Silicon, the Metal distribution enables GPU-accelerated AI/ML inference.

Included Features:

• All default features
• Metal GPU acceleration for model inference

Local Build:

CUDA (Linux)

For Linux systems with NVIDIA GPUs, CUDA distributions enable GPU-accelerated AI/ML inference. Multiple CUDA compute capability versions are available.

:::note CUDA distributions are available with the Spice Cloud Platform and Spice.ai Enterprise (opens in a new tab). Open source users can build locally for development and testing. :::

Included Features:

• All default features
• CUDA GPU acceleration for model inference

Supported Compute Capabilities:

• 80 (A100, A30)
• 86 (RTX 30xx, A40, A10)
• 87 (Jetson Orin)
• 89 (RTX 40xx, L40, L4)
• 90 (H100, H200)

Local Build:

NFS Distribution

The NFS (Network File System) distribution adds support for the NFS data connector, enabling federated queries against data stored on NFS exports.

:::note The SMB data connector is not part of this variant — smb is in the default spiced feature set, so SMB shares are queryable from the standard distribution. Only the NFS connector is feature-gated. :::

:::note The NFS connector is available with Spice.ai Enterprise (opens in a new tab). Open source users can build locally for development and testing. :::

Included Features:

• All default features
• NFS data connector

Local Build:

Allocator Variants

Different memory allocators can significantly impact performance depending on workload characteristics.

:::note Allocator variants are available with the Spice Cloud Platform and Spice.ai Enterprise (opens in a new tab). Open source users can build locally for development and testing. :::

snmalloc (Default)

The default allocator, optimized for concurrent workloads.

jemalloc

Alternative allocator that may perform better for certain memory allocation patterns.

Heap profiling: the shipped jemalloc build has the heap profiler compiled out, so _RJEM_MALLOC_CONF=prof:true is silently ignored. Build with the alloc-jemalloc-profiling feature to turn it on — it implies alloc-jemalloc and adds jemalloc's --enable-prof. jemalloc is built under the _rjem_ prefix, so it reads _RJEM_MALLOC_CONF, not MALLOC_CONF:

A profiling build reports its allocator as jemalloc-profiling rather than jemalloc at startup. In a crash report, the build is identified by the alloc-jemalloc-profiling entry in SPICED_BUILD_FEATURES.

mimalloc

Microsoft's mimalloc allocator, designed for performance and security.

System Allocator

Uses the system's default allocator (glibc malloc on Linux).

Platform Support

Platform	Default	Data	NFS	Metal	CUDA
Linux x86_64	✅	Nightly	Enterprise only	❌	Cloud/Enterprise
Linux aarch64	✅	Nightly	Enterprise only	❌	❌
macOS aarch64 (Apple Silicon)	✅	Nightly	Enterprise only	✅	❌
Windows (WSL)	✅	Nightly	Enterprise only	❌	Cloud/Enterprise
Windows (Native)	❌	Enterprise only	Enterprise only	❌	Enterprise only

:::note Native Windows support for the Spice runtime is available with the Spice Cloud Platform and Spice.ai Enterprise (opens in a new tab). Open source users on Windows should use Windows Subsystem for Linux (WSL). :::

Choosing a Distribution

Use Case	Recommended Distribution
General purpose with AI capabilities	Default
Data federation only, minimal footprint	Data (nightly)
NFS exports	NFS
macOS with GPU acceleration	Metal
Linux with NVIDIA GPU	CUDA
Memory allocation benchmarking	Allocator variants

Additional Connectors

Four data connectors sit outside the spiced default feature set, so the published open source images do not include them. Each is available in the Spice.ai Enterprise (opens in a new tab) distributions (see Distribution Availability (opens in a new tab) for Spice Cloud coverage), and open source users can build each locally for development and testing:

Connector	Cargo feature	Local build
ODBC	odbc	make install-odbc
ScyllaDB	scylladb	make install-scylladb
NFS	nfs	make install-nfs
Elasticsearch	elasticsearch	SPICED_NON_DEFAULT_FEATURES="elasticsearch" make install

For ScyllaDB specifically, a Spicepod using from: scylladb: on a build without the scylladb feature fails to load with an error naming the Cargo feature to build and linking the Enterprise distributions — the dataset is not silently skipped.

Platform-Specific Notes

Linux arm64

• FP16 (FEAT_FP16) is required because the gemm matrix multiplication library (used by the Candle ML framework) contains half-precision ARM inline assembly that requires the fullfp16 CPU feature. This is supported on AWS Graviton2+, Ampere Altra, Apple M-series (via Linux VM), and most ARMv8.2-A+ processors.
• lld is required as the linker because the spiced debug binary is large enough to exceed GNU ld's ±128 MiB branch range for R_AARCH64_CALL26 relocations. lld automatically inserts range extension thunks.
• Install prerequisites on Ubuntu/Debian: sudo apt-get install -y clang lld

Linux x86_64

• Release builds target AVX2+ for optimized SIMD performance, covering Intel Haswell (2013+) and AMD Excavator (2015+) processors, including all current AWS x86_64 instance families (C6/C7/C8).

Building Custom Distributions

Custom distributions with specific feature combinations can be built:

See the project Makefile (opens in a new tab) for all available build targets and options.