Huawei is gearing up to release Pangu Ultra—a 135B dense Transformer trained on 13.2T tokens with 8,192 Ascend NPUs. Despite fewer parameters, it matches DeepSeek-R1 in performance.
No full tech report or model yet. Check it out: https://
arxiv.org/abs/2504.07866
@jiqizhixin
-
Huawei Pangu Ultra 135B Matches DeepSeek-R1 Performance
By
–
-
DDT: Decoupled Diffusion Transformer for Image Generation
By
–
DDT: Decoupled Diffusion Transformer
Hugging Face：
https://
huggingface.co/papers/2504.05
741
…
Paper:
https://
arxiv.org/pdf/2504.05741
Code：
https://
github.com/MCG-NJU/DDT -
Decoupled Transformer Improves Diffusion Model Convergence
By
–
Can a decoupled encoder-decoder Transformer bring faster convergence and better sample quality to diffusion models?
Check out Decoupled Diffusion Transformer, today's #1 paper on Hugging Face!
From Nanjing University & ByteDance—worth a deep dive. -
Scaling Laws for Native Multimodal Models
By
–
Paper：Scaling Laws for Native Multimodal Models
Link: https://
arxiv.org/pdf/2504.07951 -
Apple Researchers Discover Scaling Laws Native Multimodal Models
By
–
Researchers from Apple just uncovered scaling laws for native multimodal models. What did they find?
– No clear advantage for late-fusion over early-fusion.
– Early-fusion performs better at small scales.
– Add MoEs, and you unlock modality-specific weights→big performance gains -
Seed-Thinking-v1.5: Advanced Reasoning Models with Reinforcement Learning
By
–
Seed-Thinking-v1.5: Advancing Superb Reasoning Models with Reinforcement Learning GitHub: https://
github.com/ByteDance-Seed
/Seed-Thinking-v1.5
…
Tech Report: https://
github.com/ByteDance-Seed
/Seed-Thinking-v1.5/blob/main/seed-thinking-v1.5.pdf
… -
Alibaba and Google Adopt MCP: Mainstream AI Protocol Shift
By
–
Yesterday, Alibaba announced they’re adopting MCP—now Google’s in the game too. Looks like MCP is going mainstream fast.
-
Alibaba Rumored to Release Qwen 3 Model Soon
By
–
Rumor has it Alibaba might drop Qwen 3 in the next few hours. Thoughts?
-
Mixture of Routers: Novel Architecture for AI Model Routing
By
–
Mixture of Routers
Paper: https://
arxiv.org/pdf/2503.23362
Code: https://
anonymous.4open.science/r/MoR-DFC6 -
Mixture of Routers: Advanced MoE Architecture for DeepSeek
By
–
MoE is powerful and a key foundation for models like DeepSeek. SUES takes it further with Mixture of Routers (MoR), applying MoE to routers! MoR uses multiple subrouters for joint selection, with a learnable main router to weight them—and it performs impressively well.