DeepSeek-V4-Flash-0731 Ships MIT-Licensed at 304B Params, Beats DeepSeek-V4-Pro, and Costs $0.14 per Million Input Tokens

DeepSeek (HuggingFace model card), corroborated by Simon Willison and Latent Space AINews·high signalXBlueskyLinkedInCopy link

DeepSeek released V4-Flash-0731 on July 31 under the MIT license with a DSpark speculative-decoding module attached, scoring Terminal Bench 2.1 82.7, Toolathlon-Verified 70.3, DSBench-FullStack 68.7, DeepSWE 54.4 and NL2Repo 54.2 — the model card states it outperforms DeepSeek-V4-Pro (Preview) 'despite its far smaller activated parameter count.' The accompanying technical report is titled 'Towards Highly Efficient Million-Token Context Intelligence,' and API pricing lands at $0.14/$0.28 per million tokens with a 98% cache discount. The combination that matters for builders: frontier-adjacent agentic scores, a million-token window, open weights you can actually redistribute commercially, and a price roughly two orders of magnitude below flagship closed models.