Story perspectives

DeepSeek Launches Cheap V4 Models, Faces IP Lawsuit

4/25/2026

Summary

DeepSeek Launches

DeepSeek released V4 Pro and Flash preview models on Friday.
V4-Pro (1.6 trillion parameters) and V4-Flash (284 billion parameters) run on Huawei Ascend AI processors.
DeepSeek priced V4-Pro at $3.48 and V4-Flash at $0.28 per million tokens, below OpenAI’s $30 and Anthropic’s $25.
SMIC shares rose 10% in Hong Kong, while Minimax and Knowledge Atlas fell over 9%.
Anthropic and OpenAI allege DeepSeek illegally distilled their models, a claim China rejects.

Summary

DeepSeek Tops Benchmark

DeepSeek's V4-Pro secured a 90.2% score on the Apex Shortlist math-coding benchmark, matching top closed-source rivals.
V4-Pro uses a mixture-of-experts design, activating only 49 billion of 1.6 trillion parameters to boost efficiency.
The model achieved a Codeforces rating of 3206 and led SimpleQA-Verified with 57.9% accuracy.
DeepSeek announced an external fundraising round to expand computing capacity and retain talent before the full model rollout.

History

Historical Context

OpenAI unveiled GPT-3 in May 2020, a 175-billion-parameter model that Microsoft licensed exclusively on 23 September 2020, sparking a race to build ever larger transformers. That scaling pressure now appears in China, where DeepSeek’s V4-Pro, built on Huawei Ascend chips, has prompted the Ministry of Industry and SMIC to weigh innovation against IP enforcement.

1 of 3

DeepSeek Launches

DeepSeek released V4 Pro and Flash preview models on Friday.
V4-Pro (1.6 trillion parameters) and V4-Flash (284 billion parameters) run on Huawei Ascend AI processors.
DeepSeek priced V4-Pro at $3.48 and V4-Flash at $0.28 per million tokens, below OpenAI’s $30 and Anthropic’s $25.
SMIC shares rose 10% in Hong Kong, while Minimax and Knowledge Atlas fell over 9%.
Anthropic and OpenAI allege DeepSeek illegally distilled their models, a claim China rejects.

1 / 3