DeepSeek has published the official DeepSeek-V4-Flash-0731 model on Hugging Face and moved its API to public beta as of July 31, 2026. This release supersedes the earlier preview version through re-post-training rather than architectural changes, delivering significant improvements in agentic and coding benchmarks while maintaining the same 284B parameter size.

  • The model features a 284B-parameter MoE architecture with 13B activated per token and includes the DSpark speculative decoding module.
  • It achieves higher scores than V4-Pro (Preview) on all published agentic benchmarks, including 82.7 on Terminal Bench 2.1 and 54.2 on NL2Repo.
  • API pricing is set at $0.14 per 1M input tokens and $0.28 per 1M output tokens, roughly a third of the cost of V4-Pro.
  • The weights are MIT-licensed and ungated, allowing self-hosting on approximately 110 GB of memory with quantization or a 4×GB300 node for full precision.

The update provides a cost-effective option for agent loops and coding tasks, offering performance that rivals more expensive models while remaining accessible to indie developers and enterprises via API or self-hosting.