The Chinese AI firm DeepSeek has unveiled its DeepSeek-V4.1-Flash model, featuring native multimodality, and announced a phased rollout of V4-Pro.
"Designed to enhance capabilities, expedite result delivery, increase throughput, and scale to larger models," the company stated.
Compactness and Performance
DeepSeek claims that V4.1-Flash is the most compact model in its new architectural family. This neural network, classified as a Mixture of Experts, boasts 552 billion parameters with a Causal Encoder-Decoder architecture: 8 billion parameters remain active at the input and 16 billion at the output.
According to DeepSeek, through new pre-training methods and more extensive reinforcement post-training, this model surpasses the performance, cost, speed, and overall execution time of DeepSeek-V4-Pro.
Source: DeepSeek.Additionally, DeepSeek reported a reduction in KV-cache requirements. The V4.1-Flash model requires only one-fourth the amount of HBM and one-eighth the SSD storage compared to the previous generation. The company believes this will lower costs in agent scenarios.
The model is now accessible via the DeepSeek API. The V4-Flash and V4-Flash-Vision-Exp have been retired, with the compatible aliases deepseek-v4-flash and deepseek-v4-flash-vision-exp temporarily redirecting requests to V4.1-Flash.
Starting at 04:00 UTC on September 14, all requests to deepseek-v4-pro will automatically reroute to V4.1-Flash at the new model's rates. This arrangement will remain in place until the launch of V4.1-Pro. The updated pricing for V4.1-Flash took effect at 04:00 UTC on September 10.
Source: DeepSeek.DeepSeek also announced plans to collaborate with the open-source community to support the inference of V4.1-Flash and explore additional deployment options.
Preparing for IPO
DeepSeek has engaged CITIC Securities to assist in preparing for its initial public offering (IPO) on the STAR Market of the Shanghai Stock Exchange, as reported by Reuters and SCMP. The largest investment bank in China will serve as one of four underwriters.
The company aims to initiate the IPO process by the end of 2026. Specific details regarding the timing, size, and valuation of DeepSeek's offering have yet to be determined.
Typically, Chinese companies undergo a preliminary preparation process with brokerage firms prior to going public. The involvement of CITIC Securities indicates that DeepSeek has advanced further in its preparations for a potential listing.
DeepSeek Could Achieve a Valuation of $75 Billion
In June, DeepSeek raised approximately $7.4 billion at a post-funding valuation exceeding $50 billion. Major investors include company founder Liang Wenfeng, Tencent, and battery manufacturer CATL.
In July, Reuters reported that a new funding round could value DeepSeek at around 500 billion yuan, equivalent to approximately $75 billion. The company intends to allocate additional funds towards computational infrastructure, model development, and retaining employees amid competition from other Chinese and American AI developers.
According to the agency, Liang Wenfeng also aims to use the raised capital to enhance the motivation of key researchers and engineers. DeepSeek has already experienced talent departures to larger competitors, including ByteDance and Xiaomi.
It is worth noting that in July, media reports indicated that TikTok's operator is shifting its focus towards artificial intelligence and is actively expanding its related infrastructure.