On this page

NVIDIA has shipped version 2.1.6 of its NIM for LLM/VLM inference containers (November 14, 2026). The release upgrades underlying backends to vLLM 0.32.0 and SGLang 0.5.22 across both model-specific and model-free NIM containers.

Five Models Graduate to First Mainline Releases

Whereas several architectures were previously offered only as standalone variant containers (such as 2.1.4-variant), NIM 2.1.6 marks the first main release for five models, alongside validating kimi-k3, minimax-m3, and qwen3.8-flash-next in nvidia/vllm-model-free-nim and glm-5.3-flash in nvidia/sglang-model-free-nim:

**SGLang Mainline Containers: **deepseek-v4.1-flash (deepseek-ai/deepseek-v4.1-flash, shifting profile prefixes from vllm- to sglang-), inkling-small (thinkingmachines/inkling-small, also moving to sglang- profiles), and muse-glimmer (meta/muse-glimmer).

**vLLM Mainline Containers: **nemotron-3.5-super-vl-120b-a12b (nvidia/nemotron-3.5-super-vl-120b-a12b) and qwen3.8-27b (qwen/qwen3.8-27b).

Cached Runtime Manifests, Optional Startup Warmup & Security Fixes

**Cached Runtime Manifests for Faster Restarts: **When NIM_MODEL_PATH uses a remote URI such as ngc:// or hf://, first startup writes a runtime manifest alongside nim_runtime_manifest.key.json inside NIM_CACHE_PATH. Subsequent restarts with matching cache, container image, and path reuse the cached manifest to avoid redundant network and authentication calls. NVIDIA recommends pinning NIM_MODEL_PATH to an immutable revision and priming the cache online before moving images into air-gapped environments.

**Optional Startup Warmup Request: **Setting NIM_STARTUP_WARMUP to 1, true, or yes (opt-in, disabled by default) instructs NIM to send a short chat completion request after the backend reports healthy and before marking the container ready.

**Security Fixes Across vLLM and SGLang Images: **NIM 2.1.6 resolves 99 unique security issues (10 critical, 30 high, 45 medium, 14 low), including 91 fixes in the vLLM image (upgrading Pillow to 13.0.0 and nginx to 1.31.7, plus addressing high-severity TLS/mTLS issue CVE-2026-97629).

Deployment Boundaries

**Container runtime update; check hardware support matrix before deploying: **Mainline promotion reflects container packaging updates rather than initial model weight releases, and specific GPU or quantization constraints (such as kimi-k3 requiring NVIDIA GB300 on the vLLM model-free container) are detailed in the official support matrix and release notes.