DeepSeek Chat's Expert Mode now runs on V4-Pro

DeepSeek has released DeepSeek-V4-Pro, its new flagship model, with 1.6 trillion total parameters and 49 billion active per token. The release was posted to Threads, and V4-Pro is the model DeepSeek Chat now serves once Expert Mode is switched on.

At a glance

• Chat users reach V4-Pro through the Expert Mode toggle in DeepSeek Chat, while teams integrating the model into their own applications reach the same release through DeepSeek's APIs.
• DeepSeek puts the model at 1.6 trillion total parameters with 49 billion active per token, meaning only a fraction of the network is engaged for each token it generates.
• The generally available build carries the label V4 Pro 0813 and appeared on August 12, 2026, after DeepSeek opened the beta of V4-Flash-0731 and promised a Pro tier to follow it.

Why it matters A flagship that ships simultaneously in the consumer product and in the API leaves little distance between the demo surface and production use, which likely matters more to practitioners than the headline parameter count. Sparse activation is what makes a 1.6 trillion parameter model plausible to serve at all, and the 49 billion figure is the one that appears to bound serving cost.

The two parameter figures describe different quantities. The 1.6 trillion total counts the full network as shipped, which sets the memory a deployment has to hold; the 49 billion active is the share that runs for any single token, and that is the figure tied to the compute spent generating each token of output.

Both entry points are live. In DeepSeek Chat, the model answers once Expert Mode is selected, and DeepSeek presents V4-Pro as the flagship of its lineup. The same release is exposed through DeepSeek's APIs for developers building it into their own products.

V4-Pro was promised when DeepSeek opened the beta of V4-Flash-0731, the earlier release in the same line. The generally available build is labelled V4 Pro 0813 and went live on August 12, 2026, carrying the Pro designation DeepSeek had signalled at that point.

What's next Pricing for V4-Pro, its context length and how it compares with V4-Flash-0731 on benchmarks are not part of the announcement. Whether the Flash line continues as a separate tier alongside the Pro build is also unstated, as is whether Expert Mode in DeepSeek Chat applies usage limits to the new model. The next signal would be documentation covering rates and quotas.