Model Early Access Builds
Per-model early access container builds shipped ahead of stable releases
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 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:X.Y.Zstable release container; the early access image is superseded.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.
Every card’s tag is click-to-copy as a full docker pull command. The coverage dots show what each build actually shipped — model builds publish container images only (no wheels, Helm charts, or crates).
v1.4.0 release
First build on the v1.4.0 line; targets the next stable release.
ImagesWheelsHelmCratesInkling recipe (main)v1.3.0 release
Container carries SGLang cherry-picks (stability, config parsing, model support) opened upstream but not yet in a released SGLang.
ImagesWheelsHelmCratesGLM-5 NVFP4 recipeDynamo changes and the M2 tool-calling fix are in release/1.3.0; the recipes run on the stock :1.3.0 containers.
ImagesWheelsHelmCratesRecipe on release branchDeepSeek-V4 Flash and Pro recipes ship in v1.3.0 on the standard TensorRT-LLM release container.
ImagesWheelsHelmCratesrecipes/deepseek-v4 (main)Four un-upstreamed vLLM patches; requires pinned flags VLLM_DISABLED_KERNELS=FlashInferFP8ScaledMMLinearKernel and --no-enable-flashinfer-autotune.
ImagesWheelsHelmCratesNemotron-3-Ultra recipeBoth container patches are in the vLLM v0.23.0 that v1.3.0 ships; the recipe runs on the stock vllm-runtime:1.3.0.
ImagesWheelsHelmCratesNemotron-3-Super recipeThe build's only container patch is in vLLM v0.23.0; the recipes run on the stock vllm-runtime:1.3.0.
ImagesWheelsHelmCratesKimi-K2.6 recipeDynamo #10132 (Cosmos3 support in the vLLM-Omni backend) is open, not merged — v1.3.0 containers cannot run Cosmos3.
ImagesWheelsHelmCratesLaunch scripts (branch)v1.2.0 release
Blackwell (B200 + GB200) preview; per-arch/CUDA tags (e.g. vllm-runtime:1.2.0-deepseek-v4-cuda13-dev.3). Superseded by the v1.3.0 recipe.
ImagesWheelsHelmCratesBlackwell preview on vLLM v0.20.0 (native DSv4 support); superseded by dev.3.
ImagesWheelsHelmCratesEarliest DSv4 preview (SGLang, B200 only); superseded by dev.2/dev.3.
ImagesWheelsHelmCratesSee Compatibility for hardware, platform, and backend feature support, and Release Artifacts for the stable release inventory.