Dynamo v1.4.0-kimi-k3-dev.1

Dynamo v1.4.0-kimi-k3-dev.1 — Release Notes

Dynamo v1.4.0-kimi-k3-dev.1 is an experimental snapshot build with early Kimi-K3 support on the Dynamo vLLM backend. It is not recommended for production and is not a QA-gated release. APIs, behavior, and defaults may change before stable support. Use it for evaluation, testing, and early feedback only.

Summary

Dynamo v1.4.0-kimi-k3-dev.1 adds experimental Kimi-K3 serving on the Dynamo vLLM backend. The published multi-architecture vLLM runtime is layered on the Kimi-K3 vLLM base image and includes frontend preprocessing, reasoning, and tool-call support for Kimi-K3 multimodal (text + image) workloads.

The release branch also includes Kimi-K3 vLLM aggregated and disaggregated Kubernetes recipes for GB300 and GB200, plus model-cache setup for the Hugging Face checkpoint.

Base Branch: main
Base Commit: 5f62ed542a20c273d38555a294dbebe0bd1bddb4 (2026-07-22)
Preview Branch: release/1.4.0-kimi-k3-dev.1
Preview Tip: ba83080ecd31c1ce918559e576d3c5bc9e092ff1
vLLM: vllm/vllm-openai:kimi-k3
NIXL: v1.1.0

Container Images

Backend	Arch	Image
vLLM (CUDA 13)	multi-arch (amd64 + arm64)	nvcr.io/nvidia/ai-dynamo/vllm-runtime:1.4.0-kimi-k3-dev.1

Backend Versions

Backend	Source	CUDA	Python	Notes
vLLM	Dynamo-built runtime on vllm/vllm-openai:kimi-k3	13.0	3.12	NIXL v1.1.0

About Kimi-K3

Kimi-K3 is Moonshot AI's hybrid-attention mixture-of-experts model with multimodal (text + image) support. This preview focuses on Dynamo + vLLM deployment recipes for agentic workloads on GB300 and GB200.

Full Changelog

vLLM Backend

• Added Kimi-K3 runtime support: Wired the Dynamo vLLM runtime to the Kimi-K3 multi-arch base image and enabled Kimi-K3 frontend preprocessing, reasoning, and tool-call handling for multimodal serving (ba83080).

Recipes

• Added Kimi-K3 vLLM deployment recipes: Added GB300 and GB200 aggregated and disaggregated agentic profiles, plus model-cache manifests, under recipes/kimi-k3 (ba83080).

Known Issues / Limitations

• This is a branch-specific snapshot build, not a QA-gated stable release.
• This preview documents vLLM recipes only.
• Profiles target multi-node GB300 (TP8) and GB200 (TP16) agentic deployments; other SKUs are not covered in this build.
• GB200 disagg frontend (recipes/kimi-k3/vllm/disagg-gb200-agentic): the Frontend passes --model-path $(MODEL_PATH), but MODEL_PATH is unset in that container, so Kubernetes leaves the literal $(MODEL_PATH) and the frontend starts with that string as its model path. Workaround: drop the --model-path / $(MODEL_PATH) lines, or use the release/1.4.0-kimi-k3-dev.1 branch tip / #12237.

---