Model Early Access Builds

Per-model early access container builds shipped ahead of stable releases

View as Markdown

See also: Release Artifacts for the stable release inventory and platform previews | Support Matrix for hardware and platform compatibility

Model early access builds do not go through QA validation. They are experimental builds intended for early testing of a specific model. They may contain bugs, require pinned runtime flags, and receive no patch support. Use a stable release container for production workloads unless the build’s GA path below says the recipe is promoted.

A model early access build packages a single model’s recipe on one runtime container, tagged X.Y.Z-<model>-dev.N and cut from a side branch ahead of that model’s launch — independently of the stable release cadence. When the model’s backend patches land upstream in the versions a stable release ships, the recipe is promoted to the plain :X.Y.Z release container and the early access image is no longer needed.

Full-platform early access builds (vX.Y.Z-dev.N, covering all runtimes, wheels, crates, and Helm charts) are platform previews and are documented under Early Access Artifacts instead.

GA path legend:

  • Promoted — the recipe runs on the stock :X.Y.Z stable release container; the early access image is superseded.
  • Dev-only — the model still requires this early access image (its patches are not yet in a stable release).
  • Recipe in GA — the model ships as a recipe on the standard release container.

Pull any image below with:

$docker pull nvcr.io/nvidia/ai-dynamo/<runtime>:<tag>

v1.4.0 release

ModelTagRuntimeShippedRecipeGA path
Inkling1.4.0-inkling-dev.1sglang-runtimeJul 17, 2026Inkling recipeDev-only (v1.4.0 line)

v1.4.0-inkling-dev.1

  • GitHub Tag: v1.4.0-inkling-dev.1
  • NGC Image: sglang-runtime:1.4.0-inkling-dev.1
  • GA path: Dev-only — first build on the v1.4.0 line; targets the next stable release.

v1.3.0 release

ModelTagRuntimeShippedRecipeGA path
GLM-5.21.3.0-glm-5.2-dev.1sglang-runtimeJul 20, 2026 (with v1.3.0)GLM-5 NVFP4 recipeDev-only — SGLang cherry-picks not upstream
MiniMax-M31.3.0-minimax-m3-dev.1vllm + sglang + tensorrtllmJun 12, 2026recipes/minimax-m3 (branch)Promoted → :1.3.0
DeepSeek-V41.3.0-deepseek-v4-dev.1tensorrtllm-runtimeJun 6, 2026recipes/deepseek-v4Recipe in v1.3.0
Nemotron-3-Ultra1.3.0-nemotron-ultra-dev.1vllm-runtimeJun 5, 2026Nemotron-3-Ultra recipeDev-only — patches not upstream; pinned flags required
Nemotron-3-Super1.3.0-nemotron-super-dev.1vllm-runtimeJun 4, 2026Nemotron-3-Super recipePromoted → :1.3.0
Kimi-K2.61.3.0-kimi-k2.6-dev.1vllm-runtimeJun 4, 2026Kimi-K2.6 recipePromoted → :1.3.0
Cosmos-31.3.0-cosmos3-dev.1vllm-runtimeJun 1, 2026examples (branch)Dev-only — Dynamo integration not merged

v1.3.0-glm-5.2-dev.1

  • GitHub Tag: v1.3.0-glm-5.2-dev.1 (ships with the v1.3.0 GA release)
  • NGC Image: sglang-runtime:1.3.0-glm-5.2-dev.1
  • Upstream status: the container carries SGLang cherry-picks (stability, config parsing, model support) opened upstream against sgl-project/sglang but not yet in a released SGLang.
  • GA path: Dev-only until the SGLang fixes land in a release that a Dynamo stable release ships.

v1.3.0-minimax-m3-dev.1

  • Branch: release/1.3.0-minimax-m3-dev.1
  • GitHub Tag: v1.3.0-minimax-m3-dev.1
  • NGC Images: vllm-runtime, sglang-runtime, and tensorrtllm-runtime at 1.3.0-minimax-m3-dev.1
  • Upstream status: Dynamo #10983 (MiniMax-M3 Dynamo changes) merged to main 2026-06-29 and is in release/1.3.0; the M2 tool-calling fix (#11554) is cherry-picked to release/1.3.0 via #11621. No backend container patches.
  • GA path: Promoted — the recipes run on the stock :1.3.0 containers.

v1.3.0-deepseek-v4-dev.1

  • GitHub Tag: v1.3.0-deepseek-v4-dev.1
  • NGC Image: tensorrtllm-runtime:1.3.0-deepseek-v4-dev.1
  • Backends: TensorRT-LLM 1.3.0rc15post1 (TRT-LLM patches over the main base)
  • GA path: Recipe in GA — DeepSeek-V4 Flash and Pro recipes ship in v1.3.0 on the standard TensorRT-LLM release container.

v1.3.0-nemotron-ultra-dev.1

  • Branch: release/1.3.0-nemotron-ultra-dev.1
  • GitHub Tag: v1.3.0-nemotron-ultra-dev.1
  • NGC Image: vllm-runtime:1.3.0-nemotron-ultra-dev.1
  • Backends: vLLM v0.22.0 with six container patches
  • Upstream status:
    • vLLM #40932 (Qwen3-VL deepstack boundary fix) — upstream, first released in vLLM v0.22.0; in v0.23.0
    • vLLM #42554 (Mamba prefix-cache PD runtime) — upstream, in vLLM v0.23.0
    • Four Ultra-specific patches (hybrid Mamba/attention hash-block KV events, MTP DS conv-state layout, SSM NIXL tail transfer, DS-tail copy i64 offsets) — not upstreamed; no vLLM PRs opened
  • GA path: Dev-only — the four un-upstreamed patches are not in the vLLM v0.23.0 that v1.3.0 ships. Requires two pinned runtime flags: VLLM_DISABLED_KERNELS=FlashInferFP8ScaledMMLinearKernel and --no-enable-flashinfer-autotune.

v1.3.0-nemotron-super-dev.1

  • Branch: release/1.3.0-nemotron-super-dev.1
  • GitHub Tag: v1.3.0-nemotron-super-dev.1
  • NGC Image: vllm-runtime:1.3.0-nemotron-super-dev.1
  • Backends: vLLM v0.22.0 with two container patches
  • Upstream status:
    • vLLM #40984 (emit KV-cache metadata) — merged 2026-05-12, first released in vLLM v0.22.0; in v0.23.0
    • vLLM #40932 (Qwen3-VL deepstack boundary fix) — upstream, first released in vLLM v0.22.0; in v0.23.0
  • GA path: Promoted — both patches are in the vLLM v0.23.0 that v1.3.0 ships; the recipe runs on the stock vllm-runtime:1.3.0.

v1.3.0-kimi-k2.6-dev.1

  • Branch: release/1.3.0-kimi-k2.6-dev.1
  • GitHub Tag: v1.3.0-kimi-k2.6-dev.1
  • NGC Image: vllm-runtime:1.3.0-kimi-k2.6-dev.1
  • Upstream status: the build’s only container patch, vLLM #40932 (Qwen3-VL deepstack boundary fix), is upstream and first released in vLLM v0.22.0; in v0.23.0 ✓; the K2.6 recipes themselves merged to main via Dynamo #10187.
  • GA path: Promoted — the recipes run on the stock vllm-runtime:1.3.0.

v1.3.0-cosmos3-dev.1

  • Branch: release/1.3.0-cosmos3-dev.1
  • GitHub Tag: v1.3.0-cosmos3-dev.1
  • NGC Image: vllm-runtime:1.3.0-cosmos3-dev.1
  • Backends: vLLM-Omni pinned to the Cosmos3 support commit
  • Upstream status:
    • vllm-omni #3454 (Cosmos3 support) — merged 2026-06-01, released in vLLM-Omni v0.23.0rc1
    • Dynamo #10132 (Cosmos3 support in the vLLM-Omni backend) — open, not merged
  • GA path: Dev-only — without #10132 the v1.3.0 containers cannot run Cosmos3, even though the vLLM-Omni side is released. Launch scripts and docs live on the build branch: docs/backends/vllm/cosmos3.md, examples/backends/vllm/launch.

v1.2.0 release

ModelTagRuntimeShippedRecipeGA path
DeepSeek-V4 preview1.2.0-deepseek-v4-dev.3vllm + sglangMay 9, 2026supersededSuperseded — recipe in v1.3.0
DeepSeek-V4 preview1.2.0-deepseek-v4-dev.2vllm + sglangMay 1, 2026supersededSuperseded — recipe in v1.3.0
DeepSeek-V4 preview1.2.0-sglang-deepseek-v4-dev.1sglangApr 25, 2026supersededSuperseded — recipe in v1.3.0

The Tag column shows the GitHub release tag. Unlike the single-image builds above, these previews publish per-arch/CUDA container tags (for example vllm-runtime:1.2.0-deepseek-v4-cuda13-dev.3) — see each build’s Container Images table below for the pullable tags.

v1.2.0-deepseek-v4-dev.3

  • Branch: release/1.2.0-deepseek-v4-dev.3
  • GitHub Tag: v1.2.0-deepseek-v4-dev.3
  • Backends: vLLM v0.20.1 (DSv4 stabilization patch over v0.20.0 native DSv4 support) | SGLang upstream lmsysorg/sglang:deepseek-v4-blackwell preview (refreshed for dev.3) | NIXL v0.10.1
  • Coverage: Partial — DeepSeek-V4-Flash and V4-Pro only. vLLM and SGLang containers are published for Blackwell (B200 plus GB200); no TensorRT-LLM container, no other component containers, no Helm charts, no wheels. Model early access build for V4 model support; not QA-gated.

Container Images

Image:TagBackendCUDAArch
vllm-runtime:1.2.0-deepseek-v4-cuda13-dev.3vLLM v0.20.1v13.0AMD64/ARM64
sglang-runtime:1.2.0-deepseek-v4-cuda12-dev.3SGLang upstream DSv4 previewv12.9AMD64
sglang-runtime:1.2.0-deepseek-v4-cuda13-dev.3SGLang upstream DSv4 previewv13.0ARM64

Python Wheels

Not published for this early access build. Use the v1.1.1 wheels or v1.1.0-dev.3 from pypi.nvidia.com.

Helm Charts

Not published for this early access build. Use v1.1.1 charts for platform install.

Rust Crates

Not shipped for early access builds.

v1.2.0-sglang-deepseek-v4-dev.1

  • Branch: release/1.2.0-sglang-deepseek-v4-dev.1
  • GitHub Tag: v1.2.0-sglang-deepseek-v4-dev.1
  • Backends: SGLang upstream lmsysorg/sglang:deepseek-v4-blackwell preview
  • Coverage: Partial — DeepSeek-V4-Flash and V4-Pro only. SGLang container only, published for Blackwell (B200). No vLLM or TensorRT-LLM containers, no other component containers, no Helm charts, no wheels. Earliest DSv4 preview build; superseded by dev.2/dev.3; not QA-gated.

Container Images

Image:TagBackendCUDAArch
sglang-runtime:1.2.0-sglang-deepseek-v4-b200-dev.1SGLang (DSv4 Blackwell preview)v12.9AMD64

Python Wheels

Not published for this early access build. Use the v1.1.1 wheels or v1.1.0-dev.3 from pypi.nvidia.com.

Helm Charts

Not published for this early access build. Use v1.1.1 charts for platform install.

Rust Crates

Not shipped for early access builds.

v1.2.0-deepseek-v4-dev.2

  • Branch: release/1.2.0-deepseek-v4-dev.2
  • GitHub Tag: v1.2.0-deepseek-v4-dev.2
  • Backends: vLLM v0.20.0 (native DeepSeek-V4 support) | SGLang upstream lmsysorg/sglang:deepseek-v4-blackwell preview | NIXL v0.10.1
  • Coverage: DeepSeek-V4-Flash and V4-Pro only. vLLM and SGLang containers are published for Blackwell. TensorRT-LLM container, other component containers, Helm charts, and wheels are not published for this tag. Model early access build for V4 model support; not QA-gated.

Container Images

Image:TagBackendCUDAArch
vllm-runtime:1.2.0-deepseek-v4-cuda13-dev.2vLLM v0.20.0v13.0AMD64/ARM64
sglang-runtime:1.2.0-deepseek-v4-cuda12-dev.2SGLang upstream DSv4 previewv12.9AMD64
sglang-runtime:1.2.0-deepseek-v4-cuda13-dev.2SGLang upstream DSv4 previewv13.0ARM64

Python Wheels

Not published for this early access build. Use the v1.1.0 wheels or v1.1.0-dev.3 from pypi.nvidia.com.

Helm Charts

Not published for this early access build. Use v1.1.0 charts for platform install.

Rust Crates

Not shipped for early access builds.