DeepSeek is training a 2‑trillion‑parameter model, The Information says
Liang Wenfeng told investors that training on Chinese chips has to work, according to summaries of the report. DeepSeek has not commented.
DeepSeek is training a model with 2 trillion parameters and has one of 8 trillion planned, The Information reported, according to three summaries of the report posted on X on Monday. DeepSeek has not commented on the figures, and the three summaries all rest on the same report.
The summaries agree on those two figures and differ on a third. Hesamation put DeepSeek's V4 Pro at 1.6 trillion parameters. The account choblin29, describing the same report, put the current V4 at 1.4 trillion. The two may be different models. Both said Moonshot's Kimi K3 runs to 2.8 trillion.
The report describes a bet on Chinese silicon. Liang Wenfeng, DeepSeek's chief executive, told investors that training on Huawei or other domestic chips is one of the company's biggest bets and that it "has to work", according to choblin29. Hesamation gave the same quotation.
DeepSeek expects Huawei chips capable of training its latest models in the fourth quarter of 2026 or the first quarter of 2027, the same summary said. It put the company's spending on new compute at about 30 billion yuan. PolymarketMoney said Liang had made domestic chips a "major priority".
The summaries also say DeepSeek is still training on Nvidia chips. The account choblin29 described those chips as likely obtained through the black market. That claim rests on a single outlet and a single summary of it. Neither DeepSeek nor Nvidia has commented.