DeepSeek-V4-Flash

Overview

DeepSeek’s efficiency-oriented MoE model in the V4 family: 284B total / 13B active parameters, 1M context, MIT weights. Official public-beta graduation as DeepSeek-V4-Flash-0731 on July 31, 2026 (API id deepseek-v4-flash).

Recent Developments

• 2026-07-31: Preview → official; architecture unchanged; improvements via re-post-training + dspark speculative decoding; native Responses API; Codex adaptation (2026-08-01-deepseek-v4-flash-0731-seawork)
• Vendor claims agent benchmarks exceed V4-Pro Preview at ⅓ Pro output price ($0.14/$0.28 per M)

Warning

Benchmark wins are vendor-reported on DeepSeek’s unreleased harness — not independently verified.

Related

• deepseek
• mixture-of-experts
• speculative-decoding
• dspark
• open-weights-agentic-models-2026
• codex
• opencode

Sources

• 2026-08-01-deepseek-v4-flash-0731-seawork
• 2026-08-01-deepseek-v4-flash-0731-official-release