DeepSeek V4 Flash 0731

• 🧠 Sparse Mixture-of-Experts: 284B total parameters, about 13B active per token
• 📏 One-million-token context window for long-document and repository-scale work
• ⚡ Efficiency-tuned V4 variant aimed at fast, high-throughput inference
• 🆕 July 2026 dated build of the V4 Flash line
• 🔧 Supports function calling, code-optimized generation and web search
• 🎯 Artificial Analysis measures 50 on its Intelligence Index at maximum effort
• 🌐 Hybrid attention design cuts long-context compute and KV cache
• 🏢 Built by DeepSeek, the Hangzhou lab backed by High-Flyer

DeepSeek

DeepSeek is a Chinese artificial intelligence company specializing in large language model development, founded in July 2023 by Liang Wenfeng. Based in Hangzhou, Zhejiang, the company is backed by High-Flyer, a prominent Chinese hedge fund also co-founded by…

DeepSeek V4 Flash 0731 is the efficiency-oriented member of DeepSeek's V4 generation: a sparse Mixture-of-Experts model with 284B total parameters and roughly 13B activated per token, paired with a one-million-token context window. It sits below the larger DeepSeek V4 Pro, which shares the same million-token context but is aimed at the heaviest knowledge and agentic workloads.

Architecturally, the V4 line departs from DeepSeek V3.2 through a hybrid attention design that combines compressed sparse attention with heavily compressed attention. DeepSeek's own model card describes this as substantially reducing per-token inference FLOPs and KV cache footprint at million-token context relative to V3.2 — the efficiency recipe that lets the Flash tier serve very long inputs at high throughput. The card also outlines a two-stage post-training pipeline: domain-specific expert models trained with supervised fine-tuning and reinforcement learning, then consolidated into a single model via on-policy distillation.

Compared with the earlier April 2026 DeepSeek V4 Flash release, the 0731 build is a dated refresh within the same family: parameter counts and context window are unchanged, so differences come from updated training rather than a new architecture. Independent evaluation from Artificial Analysis places the reasoning, maximum-effort configuration at 50 on its Intelligence Index.

Catalog capabilities include reasoning, code optimization, function calling and web search, suiting long-context agents, repository-scale coding and high-volume batch pipelines. An encrypted-serving variant is listed separately as DeepSeek V4 Flash.