DeepSeek shipped V4 Flash 0731 into public beta on July 31, 2026 — a post-training refresh of the April V4 Flash preview. The prices stay at $0.14/M input and $0.28/M output, and the weights are public on Hugging Face under the MIT License.
What’s New in 0731
The 0731 build reruns post-training on the same architecture and size as the April preview. The hosted API now serves DeepSeek-V4-Flash-0731, while the April checkpoint remains separate. DeepSeek also announced a future 2x peak pricing schedule — meaning today’s $0.14/$0.28 may not last.
Specs and Performance
V4 Flash is an efficiency-optimized Mixture-of-Experts with 284B total parameters and ~13B activated — 1M-token context, fast hybrid-attention reasoning, and output speeds around 100 tokens/sec on hosted inference. Coding is nearly on par with V4 Pro (within ~1.6 points on SWE-bench); the big gap is in complex agentic tasks and factual recall.
Pricing Context
At $0.14/$0.28, V4 Flash undercuts GPT-5.6 Luna ($0.20/$1.20) by 4.3x on output and is roughly 34x cheaper than GPT-5.6 Sol on output. Cached input drops to $0.003/M. For most developer workloads — chatbots, RAG, code completion — it is currently the cheapest near-frontier option with open weights.
Caveats
DeepSeek historically suffers availability issues during peak demand. SimpleQA-style factual recall trails Western models, and the 2x peak pricing change is coming. For mission-critical production, a routing layer with fallbacks is recommended.
Migration Notes
The legacy deepseek-chat and deepseek-reasoner aliases were retired July 24, 2026. New integrations should use the V4 model IDs directly. Self-hosting requires ~160GB of weights (2x H100 or equivalent).
