Expired50% confidenceFactExact time
Previously, Inception, Google Gemini, and the diffusion version of Gemma all attempted to train text diffusion models from scratch, and NVIDIA's modeling approach differs fundamentally from its predecessors.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/20/2026
First Seen
Valid until: 10/18/2026(expired)
Sources
Related Claims
VerifiedGoogle disclosed in the Gemini technical report that its model was exposed to multimodal data during the pre-training phase, rather than first training a language model and then performing multimodal fine-tuning.73% similarUnverifiedThe Gemini Interactions API was released after Gemini's last training run, causing the model to frequently generate code using the older Gemini 2.0.69% similarUnverifiedGemini mixes text, image, audio, and video data as input from the pre-training stage to achieve native multimodal training.67% similarUnverifiedGoogle DeepMind在Gemini技术报告中明确指出Gemini系列从设计之初就是多模态原生的，而非在语言模型基础上后期添加视觉能力66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/567286API
curl https://kongchang.com/api/v1/knowledge/claims/567286MCP
get_claim(id=567286)