DeepSeek released a preview of its V4 model family, pairing a large Pro model with a smaller Flash variant and making model weights available publicly. The company also updated its application programming interface and said a one-million-token context window would become standard across its official services.
DeepSeek described V4-Pro as a mixture-of-experts model with 1.6 trillion total parameters and 49 billion active parameters during inference. V4-Flash has 284 billion total parameters and activates 13 billion. The company positioned Pro as the higher-capability option and Flash as the faster, lower-cost choice. Both models support thinking and non-thinking modes, according to the release announcement.
The models use token-level compression and DeepSeek Sparse Attention, or DSA, which the developer said is designed to reduce computation and memory demands for long-context workloads. DeepSeek presented the one-million-token capacity as a default rather than a specialized tier, covering its chat products and API services. It did not provide independent verification of the efficiency or performance claims in the supplied announcement.
DeepSeek also claimed that V4-Pro set leading results among open models on agentic coding benchmarks and performed strongly in mathematics, science, technology, engineering and coding evaluations. The company said the model’s broader knowledge trailed only Gemini 3.1 Pro in its comparisons. It said V4-Flash approached Pro’s reasoning capability and matched it on simpler agent tasks. Those statements are vendor-reported benchmark claims and should not be read as independent model rankings.
Developers can access the models by keeping the existing DeepSeek API base address and changing the model identifier to `deepseek-v4-pro` or `deepseek-v4-flash`. The release says the service supports interfaces compatible with both OpenAI Chat Completions and Anthropic APIs. DeepSeek also listed integrations with Claude Code, OpenClaw and OpenCode, and said its own teams were using V4 for agentic coding work.
The transition carries a deadline for users of earlier endpoints. DeepSeek said `deepseek-chat` and `deepseek-reasoner` would be retired after 15:59 UTC on July 24, 2026. At the time of the announcement, those identifiers were already routing to V4-Flash in non-thinking and thinking configurations, respectively. Existing applications therefore had a three-month migration window to select explicit V4 model names and assess whether prompts, response handling and usage controls needed adjustment before the older identifiers disappeared.
The preview combines open weights with hosted access rather than limiting V4 to one distribution route. For prospective adopters, the central questions are how the models perform in independently reproduced tests, what the million-token mode costs under sustained workloads, and whether application behavior remains stable as the preview develops.
