DeepSeek opens public beta for V4-Flash API as coding benchmark improves

DeepSeek has now also released V4-Flash-0731 weights on Hugging Face under an MIT license, allowing deployment, fine-tuning and commercial use beyond the company’s official API.

26d ago

Fact Check

All claim components are confirmed by primary sources. The DeepSeek official API update log and the @deepseek_ai X account confirm the July 31, 2026 V4-Flash API public beta with strengthened programming, tool calling and long-task/agent execution. The benchmark operator Arena.ai (@arena) directly confirms the exact figures: DeepSeek-V4-Flash-High scored 1586 in the Frontend Code Arena, ranked #7 overall (and #3 open-weight). Cryptobriefing corroborates the same numbers. There is no conflicting evidence.

Summary

DeepSeek launched the public beta of its formal-version V4-Flash API, saying the release improves programming, tool calling and long-task execution while keeping the same model architecture and parameter count, with updates limited to post-training and interface changes. The company said several Agent benchmark results significantly exceeded V4-Pro-Preview. DeepSeek has now also released the official V4-Flash-0731 weights on Hugging Face under the MIT license, allowing developers to deploy, fine-tune and use the model commercially without relying only on DeepSeek’s official API. A monitoring post said the newly post-trained version outperformed V4-Pro-Preview in multiple Agent tests. Separately, DeepSeek-V4-Flash-High scored 1586 on arena.ai’s Frontend Code Arena on July 31, 2026, ranking No. 7 overall and No. 3 among open-weight models, a 154-point improvement from DeepSeek V4-Pro-Preview.

Terms & Concepts

tool calling: A model capability that allows an AI system to invoke external tools or functions to complete tasks.
Mixture-of-Experts architecture: Model design that activates only selected subnetworks for each task.
open-weight models: AI models whose trained weights are publicly available or accessible.