LLMgram · AI News · 2026-07-23

NVIDIA ships 4-step Cosmos 3 Super models up to 25x faster

NVIDIA AI announced optimized 4-step Cosmos 3 Super models that generate images and video up to 25x faster than the originals while keeping top-tier quality. The company says they currently rank number one for image-to-video and number two for text-to-image on Artificial Analysis among open-weight systems.

Key takeaway

Efficiency-tuned open world models are now competing on latency as hard as fidelity, which changes what real-time creative and simulation apps can ship on commodity inference budgets.

Context

NVIDIA positioned the new 4-step Cosmos 3 Super variants as drop-in speed upgrades over prior Cosmos 3 Super releases, targeting image and video generation without trading away competitive quality claims.

For builders, a claimed order-of-magnitude inference cut at top open-weight leaderboard slots matters more than another incremental checkpoint: it lowers the cost of interactive video loops, preview pipelines, and agentic media tools that previously stalled on latency.

Numbers to know

  • 25xClaimed max speedup versus original Cosmos 3 Super generation
  • 4-stepInference step count for the optimized Super models
  • #1Reported Artificial Analysis open-weight image-to-video rank
  • #2Reported Artificial Analysis open-weight text-to-image rank
LLMgram editorial selection and synthesis · @llmgram. LLMgram is not the original publisher of this information.
Continue on LLMgram: Open in AI Signal →
Original reporting: The new 4-step Cosmos 3 Super models generate images and video up to 25x faster than the originals, and still rank among the best open-weight models on @ArtificialAnlys. 🥇 #1 for…