DeepSeek is set to raise more than 80 billion yuan, roughly $12 billion, in a second external funding round that is expected to close in October, according to people familiar with the fundraising. Tencent and battery manufacturer CATL are reported to be among the largest participants.
The Chinese AI developer began the round in July with a target valuation of 500 billion yuan, about $74 billion. Reuters reported that committed capital had exceeded an initial fundraising target of 50 billion yuan, citing a person with knowledge of the matter. A second person also said the round was expected to finish in October.
The funding has not been announced by DeepSeek. The people discussing the closing timetable declined to be identified because the information was confidential, and the company did not immediately respond to a request for comment.
Bloomberg, in reporting also carried by TNW, identified Tencent and CATL as making some of the largest commitments. That account said the final total could approach 100 billion yuan. Those figures describe a fundraising still in progress, rather than a completed transaction.
DeepSeek raised about $7.4 billion in its first external round in June. The new round’s reported valuation target is distinct from the amount being raised. Tencent operates WeChat and a cloud business; CATL is a major supplier of electric-vehicle batteries.
Separately, DeepSeek has hired CITIC Securities to prepare for a potential listing on Shanghai’s STAR Market. The timing, valuation and amount to be raised in that possible initial public offering remain undecided.
The laboratory originated from High-Flyer, the quantitative hedge fund established by founder Liang Wenfeng. It gained international attention with its R1 model in early 2025 and has continued releasing models that developers can download and run themselves.
DeepSeek’s latest release, V4.1-Flash, accepts images and text and generates text, with a context window of up to one million tokens. Its model card describes a mixture-of-experts architecture with 552 billion backbone parameters and an additional conditional-memory component. The company publishes both the model weights and repository under the MIT license.
The design reduces the memory used to retain information during generation. DeepSeek says its global key-value cache occupies 890 bytes per token, roughly a quarter of the corresponding cache in V4-Flash. The release includes inference code and tools for reproducing its DeepSWE coding benchmark results.
DeepSeek has also released programming libraries for Huawei’s Ascend chips. The tools cover matrix calculations and communication between the experts in a mixture-of-experts model, with interfaces aligned to libraries it developed for NVIDIA hardware.
Sources: Reuters, TNW, DeepSeek
–
By the Control Plane Editorial Team