DeepSeek plans to pretrain its next-generation model, DeepSeek V5, on Huawei's Ascend 950 chips, The Paper and STAR Market Daily reported on November 3. If confirmed, V5 would be the first DeepSeek model trained natively on Ascend silicon — until now, Ascend chips mainly handled inference within DeepSeek's stack.
STAR Market Daily cited multiple supply-chain sources on the plan, while The Paper confirmed the direction and added chip details: the Ascend 950 series comes in 950PR and 950DT variants, both built on the third-generation Da Vinci architecture and aimed at training and inference for large models.
Pretraining is a far tougher test than inference. It requires clusters to run at scale for weeks, stressing interconnect bandwidth, fault tolerance and the software stack. Placing a flagship model's pretraining on one chip route is, in effect, an endorsement of that route's maturity. Earlier this week, Meituan said its LongCat-2.5 was trained entirely on domestic compute; chip choices at top labs are turning from isolated trials into a route-level contest.
The training shift comes alongside rapid iteration on DeepSeek's agent stack. On November 2, the team showed Creator Mode in DeepSeek Harness: users describe a tool in plain language and the system turns it into a plug-and-play plugin — a design workbench, a sandbox game or a 3D model visualizer — with seven bilingual prompt demos published. Harness is the software framework that lets the model operate as an agent, managing and executing complex tasks.
On October 31, the same team published a technical article on a "Super-agent" design, using a two-layer architecture, event-driven scheduling, compute-storage separation and an external watchdog to take over session management and context sync; internal instances have autonomously found and fixed software defects with no human trigger.
Neither Huawei nor DeepSeek has officially confirmed the training arrangement; details await follow-up disclosure. If V5 does pretrain on Ascend, the contest in domestic AI compute moves from inference replacement to head-on training competition.