DeepSeek previews DeepSeek V4 Flash and V4 Pro AI models

In short: DeepSeek has released preview versions of two new AI text models, called DeepSeek V4 Flash and DeepSeek V4 Pro.

What happened

DeepSeek, a Chinese AI lab, published two preview models in its DeepSeek V4 family: V4 Flash and V4 Pro. These are large language models, which are AI systems that read and write text.

DeepSeek says both models can handle up to 1 million tokens of context. Tokens are small chunks of text, so this is like being able to keep a very large book, or a big code project, “in mind” while answering questions.

The company also said both models use a “mixture-of-experts” design (like a team where only the right specialists speak up for each job). DeepSeek says this helps lower the cost to run the model.

DeepSeek V4 Pro is listed at 1.6 trillion total parameters, with 49 billion active at a time. Parameters are the stored “knobs” inside a model that shape how it responds. The smaller V4 Flash has 284 billion total parameters, with 13 billion active.

DeepSeek claims the models perform better than its prior V3.2 release on reasoning tests, but still lag behind top models on some knowledge tests. The models are text-only, meaning they do not directly work with images, audio, or video.

DeepSeek also published pricing that is lower than many leading competitors. V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens, while V4 Pro costs $0.145 per million input tokens and $3.48 per million output tokens.

Why it matters

Cheaper models with very large context limits can make AI features more affordable for apps that summarize long documents, search through big files, or help write and review software. At the same time, these are previews, and DeepSeek faces ongoing scrutiny and accusations from some US AI companies about copying their models.

Source: TechCrunch AI