INCREDIBLE STUFF INCOMING
The new ~500B parameter model shows competitive performance in leaked benchmarks against leading Chinese LLMs.
NVIDIA is entering the high-end large language model arena with its Nemotron 3 Ultra Base, a model reportedly containing approximately 500 billion parameters. The news, which surfaced via a viral social media post, centers on leaked benchmark results that pit this new model against two of China's most advanced AI systems: Kimi K2 from Moonshot AI and the latest GLM model from Zhipu AI. The preliminary data suggests Nemotron 3 Ultra Base's performance is 'looking good,' indicating it can hold its own against these established regional leaders. This move signals NVIDIA's ambition to compete not just as a hardware provider but as a creator of frontier AI software.
While specific benchmark scores and tasks remain undisclosed, the comparison to Kimi K2 and GLM is significant. Kimi K2 is renowned for its exceptionally long context window, while GLM models are core to China's domestic AI ecosystem. A competitive showing from NVIDIA's model would challenge the current narrative of regional specialization and could influence global AI development strategies. The 'Ultra Base' designation also hints at a family of models, suggesting more specialized or instruction-tuned variants may follow, expanding NVIDIA's direct offerings to developers and enterprises.
- NVIDIA's new Nemotron 3 Ultra Base model is estimated at ~500 billion parameters.
- Leaked benchmarks show it is competitive with top Chinese models Kimi K2 and GLM.
- This marks NVIDIA's expansion from AI hardware into creating frontier-scale software models.
Why It Matters
It challenges regional AI dominance and positions NVIDIA as a full-stack AI contender, impacting global competitive dynamics.