DeepSeek-V4 Preview launches with open weights and API access

Chinese AI company DeepSeek caused quite a stir last year as it climbed to the top of the App Store charts and beat out incumbents like ChatGPT. That was over a year ago and now the company has launched a preview of its latest and much more capable model dubbed DeepSeek-V4.

If you have a DeepSeek account, you can try out the new model here: chat.deepseek.com/. There are two versions available, Expert and Instant. Expert is a massive 1.6 trillion parameter model, Instant is a more reasonable (but still huge) 284 billion parameter model.

DeepSeek-V4 Pro and Flash models

And here’s the most interesting part – this is an open-weights model, meaning that you can download it from Hugging Face and run it on your own hardware at home. Well, you could theoretically – you would need some serious hardware to actually run these models.

Now, the Expert (Pro) model has 49 billion active parameters and Instant (Flash) has only 13 billion active parameters – these are the parameters that actually need to fit in VRAM. However, shuffling parameters between VRAM and system memory slows down token generation. The good thing about open-weight models is that the community can create quantized and distilled models that can actually run on consumer-level hardware.

Anyway, the company says that DeepSeek-V4-Flash, the model that powers Instant mode, has reasoning capabilities close to that of V4-Pro and performs just as well on simple agentic tasks. Both models have a 1 million token context window.

DeepSeek-V4-Pro, which is behind the Expert mode, has the best reasoning skills for math, STEM and coding tasks among open-weight models and rivals the performance of closed-source models, according to DeepSeek testing. V4-Pro also beats most AI models in terms of world knowledge, only losing to Google’s Gemini-3.1-Pro, says the company.

DeepSeek-V4-Pro versus its major competitors

DeepSeek-V4 can be used with popular tools like Claude Code, OpenClaw and OpenCode. If you can’t run the model at home, you can use the company’s API. Here’s the pricing:

DeepSeek-V4 Pro and Flash API pricing

Source