Skip to content

Release Notes

2.1.0

Image: nvcr.io/nim/nvidia/cosmos3:2.1.0.

New Models and Features

  • Edge Reasoner (NIM_MODEL_VARIANT=edge): BF16 and FP8 profiles with a 12-GiB per-device floor and no DFlash draft.
  • Edge-DROID policy (NIM_MODEL_VARIANT=edge-droid): a BF16 Generator policy with action-only output. It is experimental because it has not been validated.
  • Streaming video sessions for the Reasoner over WebSocket and REST (NIM_ENABLE_STREAMING=true). Refer to Streaming.
  • NVIDIA Dynamo deployment for the Reasoner, including REST streaming. Refer to Dynamo Deployment.
  • Reasoner optimization presets (NIM_OPTIMIZATION_PRESET=TTFT or TPOT), which are experimental. Refer to Configuration.
  • Generator Hugging Face checkpoints: NIM_MODEL_PATH accepts hf://owner/repository[:revision] for both runtimes.
  • Transfer control weights: each control object accepts a relative control_weight. Refer to Transfer.
  • AVI video input for both runtimes.

Behavior Changes

  • Transfer guidance: with control and text guidance both active, the control term is no longer also multiplied by guidance_scale. The effective edge-family coefficient is 1.5 instead of 4.5, so combined-guidance outputs differ from 2.0.0 baselines. To approximate the 2.0.0 strength, multiply the previous control_guidance by guidance_scale, up to the 10.0 maximum. Refer to Control Guidance Behavior.
  • Generator performance defaults: SeaCache is on for every Generator variant except Edge-DROID. SageAttention 2.2 is the default on compute capability 8.0 through 9.x without sequence parallelism. The channels-last variational autoencoder (VAE) decode path is the default. If you notice a quality issue in a specific workload, try NIM_CACHE_BACKEND=none or NIM_VAE_FAST_PATH=lossless. Refer to Configuration.
  • System-memory floors are enforced: each profile carries its measured system-memory floor instead of a 16-GiB admission tag, and FP8 offload floors are lower. Refer to the Support Matrix.
  • Unified memory: on NVIDIA DGX Spark and Jetson AGX Thor, 70% of the shared pool counts toward profile floors, and the Reasoner defaults to NIM_GPU_MEMORY_UTILIZATION=0.70. Super Generator profiles no longer fit these systems.
  • Reasoner headroom: NIM_GPU_MEMORY_HEADROOM_GIB defaults to 4 GiB on discrete GPUs.
  • Validation: flow_shift must be greater than 0.

Renamed or Removed Variables

2.0.0 2.1.0
NIM_OFFLOAD_TEXT_GUARDRAIL text_guard_offload key in NIM_TAGS_SELECTOR (resident or offloaded)
NIM_OFFLOAD_VIDEO_GUARDRAIL video_guard_offload key in NIM_TAGS_SELECTOR (resident or offloaded)
NIM_UNIFIED_MEMORY_HOST_RESERVE_GIB Removed; unified-memory systems use a fixed 70% share
NIM_QWEN3GUARD_MAX_GPU_MEMORY_GB Removed

Known Issues

  • Edge Reasoner streaming: With NIM_ENABLE_STREAMING=true, the default NIM_STREAMING_SPEC_METHOD=dflash fails for Edge, which does not support DFlash. Set NIM_STREAMING_SPEC_METHOD=none. Refer to Deploy Edge Models.

2.0.0

This is the first release of the unified Cosmos3 NIM, published as nvcr.io/nim/nvidia/cosmos3:2.0.0. One container image includes both runtime choices:

  • Generator serves image and video generation, Action, and Transfer through POST /v1/infer.
  • Reasoner serves OpenAI-compatible image and video understanding through Chat Completions and the Responses API.

Refer to Deployment for launch instructions.