This document provides a comprehensive inventory of all Dynamo release artifacts including container images, Python wheels, Helm charts, and Rust crates.
See also: Support Matrix for hardware and platform compatibility | Feature Matrix for backend feature support
Release history in this document begins at v0.6.0.
Experimental: v1.2.0-deepseek-v4-dev.2 (DeepSeek-V4-Flash / V4-Pro on Blackwell, vLLM + SGLang containers only) is available as an experimental preview. Tagged Pre-Releases and experimental builds are listed under Pre-Release Artifacts.
We recommend using the TensorRT-LLM NGC container instead of the ai-dynamo[trtllm] wheel. See the NGC container collection for supported images.
The dynamo-crds Helm chart is deprecated as of v1.0.0; CRDs are now managed by the Dynamo Operator. The dynamo-graph Helm chart is deprecated as of v0.9.0.
For detailed run instructions, see the backend-specific guides: vLLM | SGLang | TensorRT-LLM
For detailed installation instructions, see the Local Quick Start in the README.
For Kubernetes deployment instructions, see the Kubernetes Installation Guide.
For API documentation, see each crate on docs.rs. To build Dynamo from source, see Building from Source.
CUDA and Driver Requirements: For detailed CUDA toolkit versions and minimum driver requirements for each container image, see the Support Matrix.
For a complete list of known issues, refer to the release notes for each version:
Each bullet is a delta to what ships on NGC / Helm / PyPI / crates.io: net-new crates, removed Helm charts, or image lines that split or appear on the registry. See the inventory tables above for full matrices.
Stable releases first (newest first). Pre-Release Git Tags (v*-dev.*, experimental tracks) are summarized below; per-tag images and wheels are spelled out in Pre-Release Artifacts.
For backend version pins, see the version-pins table above and the GitHub Releases table below.
Stable Releases
v0.5.10.post1 (NIXL v1.0.1), TRT-LLM v1.3.0rc11 (NIXL v0.10.1), vLLM v0.19.0 (NIXL v0.10.1).dynamo-planner image on NGC for Profiler jobs and Planner pods; worker and runtime images no longer bundle Planner (artifact boundary change, not a new engine capability). Crates: First 1.y.z publication on crates.io for dynamo-protocols (multi-protocol types; dynamo-async-openai remains deprecated with final release 1.0.2).snapshot-agent, EFA variants for vLLM and TRT-LLM (AMD64 only). Crates: First publish of dynamo-mocker, dynamo-kv-router. Helm: Added snapshot (preview); dropped deprecated dynamo-crds from the publish stream (CRDs owned by the Operator).dynamo-tokens. Helm: Dropped deprecated dynamo-graph from the publish stream.dynamo-frontend, CUDA 13 variants for vLLM and SGLang. Crates: First publish of dynamo-memory, dynamo-config.Dynamo Nightlies
ai-dynamo and ai-dynamo-runtime — nightly builds from main publish wheels tagged *.devYYYYMMDD. Install with pip or uv using --pre and the same NVIDIA extra-index pattern as Pre-Release Artifacts.* *.devYYYYMMDD versioning for nightly main wheels began Apr 24, 2026.
Pre-Release and Experimental Git Tags
vllm-runtime:*-deepseek-v4-cuda13-dev.2, sglang-runtime:*-deepseek-v4-cuda12-dev.2, sglang-runtime:*-deepseek-v4-cuda13-dev.2. Helm / PyPI: Not published for this tag (see Pre-Release Artifacts).tensorrtllm-runtime:1.1.0-dev.3. Wheels: ai-dynamo, ai-dynamo-runtime on pypi.nvidia.com (see below).sglang-runtime:1.1.0-dev.2, tensorrtllm-runtime:1.1.0-dev.2. Wheels: ai-dynamo, ai-dynamo-runtime on pypi.nvidia.com (see below).dynamo-frontend, kubernetes-operator, snapshot-agent. Wheels: ai-dynamo, ai-dynamo-runtime on pypi.nvidia.com. Helm: dynamo-platform, snapshot at 1.1.0-dev.1 (see below).Helm-Only Patches
dynamo-platform Helm chart only (operator image tag correction).Backend-Only Patch Trains
These crates use repository https://github.com/ai-dynamo/dynamo.git. The table lists each crate’s first non-placeholder publication on crates.io (excluding reservation uploads named 0.0.0-prerelease.0). Dates are from the crates.io registry index.
dynamo-async-openai is deprecated; 1.0.2 is its final crates.io release. Use dynamo-protocols for new dependencies (crate).
dynamo-tokenizers exists only in the Dynamo workspace and is not published on crates.io.
NGC Collection: ai-dynamo
To access a specific version, append
?version=TAGto the container URL:https://catalog.ngc.nvidia.com/orgs/nvidia/teams/ai-dynamo/containers/{container}?version={tag}
PyPI: ai-dynamo | ai-dynamo-runtime | kvbm
To access a specific version:
https://pypi.org/project/{package}/{version}/
NGC Helm Registry: ai-dynamo
Direct download:
https://helm.ngc.nvidia.com/nvidia/ai-dynamo/charts/{chart}-{version}.tgz
The dynamo-crds Helm chart is deprecated as of v1.0.0. CRDs are now managed by the Dynamo Operator.
dynamo-graph Helm chart is deprecated as of v0.9.0.crates.io: dynamo-runtime | dynamo-llm | dynamo-protocols | dynamo-async-openai (deprecated) | dynamo-parsers | dynamo-memory | dynamo-config | dynamo-tokens
To access a specific version:
https://crates.io/crates/{crate}/{version}
On crates.io, dynamo-protocols lists 1.1.0 as its first installable release (placeholder reservation 0.0.0-prerelease.0 omitted here like other 0.0.0-prerelease.* uploads). Earlier semver lines for the OpenAI-compatible client shipped under dynamo-async-openai — see #### dynamo-async-openai (crate) below.
Deprecated. Prefer dynamo-protocols. This crate remains published on crates.io for manifests pinned to the old package name.
Pre-Release artifacts do not go through QA validation. Pre-release versions are experimental previews intended for early testing and feedback. They may contain bugs, breaking changes, or incomplete features. Use stable releases for production workloads.
Pre-Release Python Wheels are published on the NVIDIA package index at pypi.nvidia.com, not on the public PyPI index. Like stable wheels, they are Linux (manylinux) builds for the Python versions in the Support Matrix; pip/uv on macOS or Windows will not find matching wheels. Install on a supported Linux host or inside a Linux container.
Install by adding that URL as an extra index and allowing pre-releases (PEP 440 dev versions):
A GitHub or container tag v1.1.0-dev.N maps to a wheel version 1.1.0.devN (for example v1.1.0-dev.2 → ==1.1.0.dev2). Optional extras such as ai-dynamo[vllm] use the same flags; pin the version you want from the sections below.
v0.20.0 (native DeepSeek-V4 support) | SGLang upstream lmsysorg/sglang:deepseek-v4-blackwell preview | NIXL v0.10.1Not published for this dev release. Use the v1.1.0 wheels or v1.1.0-dev.3 from pypi.nvidia.com.
Not published for this dev release. Use v1.1.0 charts for platform install.
Not shipped for pre-release versions.
v0.5.10.post1 | TensorRT-LLM v1.3.0rc11 | vLLM v0.19.0 | NIXL v0.10.1ai-dynamo and ai-dynamo-runtime wheels on pypi.nvidia.com. SGLang and vLLM containers, component containers (dynamo-frontend, dynamo-planner, kubernetes-operator, snapshot-agent), kvbm wheel, and Helm charts are not published for this tag.Available from pypi.nvidia.com (pre-release index):
kvbm==1.1.0.dev3 is not yet published.
Not published for this dev release. Use the latest stable (v1.1.0) for platform install.
Not shipped for pre-release versions.
v0.5.9 | TensorRT-LLM v1.3.0rc9 | vLLM v0.19.0 | NIXL v0.10.1ai-dynamo and ai-dynamo-runtime wheels on pypi.nvidia.com. vLLM runtime container, component containers (dynamo-frontend, dynamo-planner, kubernetes-operator, snapshot-agent), kvbm wheel, and Helm charts are not published for this tag.Available from pypi.nvidia.com (pre-release index):
Not published for this dev release. Use the latest stable (v1.1.0) for platform install.
Not shipped for pre-release versions.
v0.5.9 | TensorRT-LLM v1.3.0rc5.post1 | vLLM v0.17.1 | NIXL v0.10.1Available from pypi.nvidia.com (pre-release index):
Not shipped for pre-release versions.