DeepSeek V4 Flash

DeepSeek V4 Flash – Efficient Sparse MoE with Hybrid Long-Context Attention

DeepSeek V4 Flash is the efficiency-tier model in DeepSeek's V4 series: a 284B total parameter, 13B active parameter Mixture-of-Experts language model sharing the same hybrid attention architecture as DeepSeek V4 Pro, positioned as the fast and economical option within the V4 family. Both V4 models support a 1 million-token context window.

Architecture

V4 Flash shares the three core architectural innovations introduced across the V4 series:

• Hybrid Attention (CSA + HCA): Combines Compressed Sparse Attention and Heavily Compressed Attention to reduce long-context inference cost at the 1M-token scale
• Manifold-Constrained Hyper-Connections (mHC): Strengthens residual connections to stabilize signal propagation while preserving model expressivity
• Mixed Precision Training: MoE expert parameters use FP4 precision; most other parameters use FP8.

Reasoning Workflow

Three configurable reasoning effort modes per request, identical in structure to V4 Pro:

• Non-think — fast responses, no chain-of-thought
• Think High — reasoning-augmented responses
• Think Max (V4-Flash-Max) — maximum reasoning effort; officially documented as achieving comparable reasoning performance to V4 Pro when given a larger thinking budget, with the noted limitation that its smaller parameter scale places it slightly behind V4 Pro on pure knowledge tasks and the most complex agentic workflows.