DeepSeek V4 Flash (Non-reasoning)

DeepSeek V4 Flash is a high-efficiency Mixture-of-Experts (MoE) language model developed by DeepSeek, designed to provide frontier-class performance with significantly reduced computational overhead. Released as part of the DeepSeek V4 series, it serves as the fast, economical counterpart to the larger V4 Pro flagship. The model features a 284B total parameter architecture but only activates approximately 13B parameters per token during inference, allowing it to achieve high throughput and low latency suitable for real-time applications and high-volume agentic workloads.

The model introduces several architectural innovations aimed at long-context efficiency and training stability. It utilizes a Hybrid Attention mechanism that combines Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). This design drastically reduces the memory footprint of the Key-Value (KV) cache, enabling the model to natively support a 1 million token context window while requiring significantly less hardware memory than previous generations. Additionally, the architecture incorporates Manifold-Constrained Hyper-Connections (mHC) to enhance signal propagation across its deep layers and improve model expressivity.

DeepSeek V4 Flash is characterized by its switchable reasoning modes: Non-think, Think High, and Think Max. The Non-reasoning (or "Non-think") mode is the default setting optimized for speed and routine tasks, providing immediate, intuitive responses without the overhead of extended internal chains of thought. For complex STEM, coding, or logical problems, the model can be switched to its "Think" modes, where it allocates more compute to internal reasoning before outputting a final answer.

As an open-weight model, DeepSeek V4 Flash is distributed under the MIT license, supporting both commercial and research use. Its training involved a massive corpus of over 32 trillion tokens, followed by a post-training pipeline featuring Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO). This process integrated domain-specific expertise into a unified model, resulting in strong performance across agentic benchmarks, coding, and general language understanding.