"DeepSeek trained R2 on Ascend" rumor is 90% fake, argues AI researcher citing model-line evidence
NunoSempere · x · 2026-09-25
Responding to reporting that DeepSeek tried to train its R2 model on Huawei Ascend chips, teortaxesTex lays out why he thinks it's a broken-telephone rumor:
- He's 90% sure no "R2" project ever existed — DeepSeek unifies model lines and serves as few checkpoints as possible, so DS-R was bound to be folded in like DS-Coder
- Why test a flagship on a new stack instead of a V3.5-lite? Zhipu trained a 9B GLM-Image on Ascends but still runs flagships on Nvidia
- The best Huawei hardware then was the weak 910C with terrible software support; the only recent large model on such hardware came from Meituan
- NunoSempere notes the "DS never trained on Huawei" conclusion likely holds, but it mainly says something about Epoch's evidence standards
Related event: Blogger disputes report that DeepSeek trained R2 on Huawei Ascend(2 posts)→
More from Infra
- UK Sovereign AI invests in Fractile to build next-gen frontier AI chips — charlieharris01 · 2026-11-11
- What happened when I moved vector retrieval into Snowflake SQL instead of orchestrating RAG outside — dima806_dima · 2026-11-11
- AI chip startup Fractile raises $720M Series C for DRAM-on-processor design — Tom_Westgarth15 · 2026-11-11
- Mobile AI Factory: agentic pipeline builds on-device SLMs that beat 5-shot frontier models — kalyan_kpl · 2026-11-11
- Vivo's Starlink direct-to-cell service hits 10M users in two weeks across Brazil — XFreeze · 2026-11-11
- Nan Ya Plastics to spend NT$11.53B on fifth glass fabric plant on AI server demand — zephyr_z9 · 2026-11-11