Unverified50% confidenceReleaseExact time
China Telecom's open-sourced Xingchen series models adopt a MoE architecture, support 256K context extendable to 512K, and are among the first large models trained entirely on Ascend 910C.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/18/2026
First Seen
Valid until: 12/17/2026
Sources
Related Entities
Related Claims
UnverifiedZhipu's model is reportedly trained on Huawei Ascend 910D chips, with the 910D targeting a benchmark against NVIDIA's H100.67% similarVerifiedHuawei's Ascend series AI chips use a proprietary DaVinci architecture; the Ascend 910 series targets high-performance training, while the Ascend 310 series focuses on edge inference scenarios.64% similarUnverifiedThe model trained a rank-256 LoRA adapter on top of the base model huihui-ai/Huihui-Qwen3.8-27B-abliterated.60% similarUnverifiedHanxu Technology has released a new architecture that keeps model weights resident in memory chips, attempting to bypass the HBM bandwidth bottleneck60% similarUnverifiedByteDance has built self-developed clusters of tens of thousands of GPUs to support large-scale model training60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930935API
curl https://kongchang.com/api/v1/knowledge/claims/930935MCP
get_claim(id=930935)