DeepSeek has officially launched DeepSeek-V4-Flash-0731, moving its API into public beta and superseding previous previews. This release isn't a new architectural design but a significantly enhanced version achieved through re-post-training, building on the existing 284-billion-parameter Mixture-of-Experts (MoE) framework. Key additions include the DSpark speculative decoding module and native support for the Responses API format, alongside adaptations for Codex, signaling a focus on improved agentic and coding capabilities.
The underlying technology is a sophisticated 284B-parameter MoE model, where only 13 billion parameters are activated per token, coupled with an impressive 1-million-token context window. Its efficiency is further bolstered by a hybrid attention mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), alongside Manifold-Constrained Hyper-Connections (mHC) replacing conventional residual connections. The gains in this specific release are attributed to refined post-training, optimizing performance without altering the core architecture.
This upgrade significantly lowers the barrier to entry for developers and researchers, especially for agentic workflows and coding tasks. DeepSeek-V4-Flash-0731 is highly accessible via API with competitive pricing, making it feasible for seed-stage startups and indie developers to run complex agent loops without needing a substantial GPU budget. While self-hosting the MIT-licensed weights demands significant memory resources, it offers flexibility for mid-size and large enterprises with robust serving clusters, enabling broad adoption and experimentation with advanced AI capabilities.