NVIDIA Cosmos, StepFun checkpoints land on Ascend

// 2h agoMODEL RELEASE

NVIDIA Cosmos, StepFun checkpoints land on Ascend

Chinese developer outlet Modelers recently published eleven technical updates regarding the Ascend ecosystem, which center on two key model releases: NVIDIA's autoregressive world model (Cosmos) and a new base model checkpoint from StepFun. These additions reflect ongoing industry efforts to optimize top-tier physical AI world models and domestic foundational LLMs for Huawei's Ascend computing platform.

// ANALYSIS

Amid high-volume developer media coverage, the real signal lies in cross-hardware optimization as global physical AI models and domestic foundation checkpoints land on Ascend. Porting NVIDIA's autoregressive Cosmos model expands physical and visual AI capabilities on domestic hardware architecture. StepFun's release of a new base checkpoint demonstrates steady software-hardware alignment for Chinese LLM developers. Cross-platform portability and CAN/Ascend stack integration remain pivotal operational priorities for frontier AI deployments.

// TAGS

nvidiacosmosstepfunascendmodel-releaseworld-modelsaichina

DISCOVERED

2h ago

2026-08-02

PUBLISHED

2h ago

2026-08-02

RELEVANCE

7/ 10

AUTHOR

AiChinaNews

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

OPEN SOURCE19m ago

DeepSeek adds Reasonix to official agent docs

Reasonix is an open-source terminal coding agent built specifically for the DeepSeek API that is now featured in DeepSeek's official documentation. Built for the command line, it features a cache-first execution loop for prefix caching efficiency, automatic tool-call repair, and single-command model switching.

MODEL37m ago

Frontis-MA1 and OpenMLE enable recursive ML self-improvement

Frontis-MA1 (35B) is an AI agent built for recursive self-improvement in machine learning engineering, released alongside the full OpenMLE stack. Post-trained on four foundational operators (Draft, Improve, Debug, Crossover), the model enables continuous learning and evolutionary task execution.

UPDATE48m ago

Mercury Agent requires confirmation for destructive system changes

Mercury Agent is an open-source AI developer agent built with a focus on permission control and task safety. Instead of blindly automating all actions and risking unintended consequences, Mercury Agent prompts users for approval before executing high-risk or destructive operations, striking a balance between automation and developer oversight.