DeepSeek has introduced V4-Flash, a mixture-of-experts model with 284 billion total parameters, 13 billion active parameters and a 1 million-token context window. The company positions it as a faster, lower-cost option for reasoning and simpler tool-using workloads.
DeepSeek has introduced V4-Flash, a large language model designed as a faster alternative to its V4-Pro model for reasoning and relatively straightforward tool-using tasks.
In its release announcement, DeepSeek describes V4-Flash as a mixture-of-experts (MoE) model with 284 billion total parameters and 13 billion active parameters per inference. MoE designs route each request through only part of a model’s parameter set, which can reduce the compute required to generate a response compared with activating all parameters at once.
DeepSeek says V4-Flash’s reasoning performance comes close to V4-Pro, while its performance on simple agent-style tasks is comparable and its responses are faster. Those are company claims rather than independently established results, but they outline the intended role for the model: handling high-volume reasoning and automation workloads where latency and cost matter.
DeepSeek’s API documentation lists the model as DeepSeek-V4-Flash-0731 and specifies a 1 million-token context window. The same documentation gives it a maximum output length of 384,000 tokens.
That context capacity could be useful for applications that need to process extensive source material, such as large codebases, long document collections, detailed research records or lengthy conversations. A large context limit alone does not guarantee reliable analysis of that material, however. Results will still depend on the prompt, the quality and structure of the supplied information, and any retrieval or tool system around the model.
The API supports both thinking and non-thinking modes, according to DeepSeek’s pricing and model documentation. This gives developers a choice between configurations intended to devote more computation to reasoning and configurations aimed at more direct responses.
DeepSeek lists V4-Flash API pricing at $0.14 per million cache-miss input tokens and $0.28 per million output tokens. The model is available through DeepSeek’s API, according to the company’s product documentation.
DeepSeek’s official Hugging Face model card also identifies V4-Flash as a 284B-total, 13B-active MoE model with a 1 million-token context capacity. It distinguishes V4-Flash from a Flash-Max configuration, saying Flash-Max can approach Pro-level reasoning when allocated a larger thinking budget.
The same model card sets limits on that positioning. DeepSeek says Flash-Max trails its higher-end offerings on knowledge-intensive tasks and most highly complex agent workflows. That caveat suggests the Flash models are not presented as universal replacements for the Pro tier, particularly for difficult tasks that require broad factual knowledge or sustained multi-step execution.
Associated Press reported DeepSeek’s claims that V4-Flash approaches V4-Pro in reasoning and matches it on simpler agent tasks. AP also cited an Omdia analyst who described the V4 benchmarks as highly competitive with U.S. rivals.
For developers, the more directly verifiable proposition is the model’s specification: a large MoE architecture, configurable reasoning behavior, a 1 million-token context window and API pricing targeted at lower-cost deployment. Independent evaluations across coding, factual accuracy, tool use and long-horizon task completion will be needed to determine how those capabilities perform in production.
A faster model in DeepSeek’s V4 lineup DeepSeek has introduced V4 Flash , a large language model designed as a faster alternative to its V4 Pro model for reasoning and relatively straightforward tool using tasks.
In its release announcement, DeepSeek describes V4 Flash as a mixture of experts (MoE) model with 284 billion total parameters and 13 billion active parameters per inference.
MoE designs route each request through only part of a model’s parameter set, which can reduce the compute required to generate a response compared with activating all parameters at once.
Continue reading
