DeepSeek Open-Sources V4-Flash-0731 with Native Responses API Support
solyarisoftware · x · 2026-08-01
DeepSeek has officially open-sourced DeepSeek-V4-Flash-0731 and launched its public beta API.
The model is a sparse MoE architecture featuring 256 routed experts (6 active per token), a 1M-token context window, and three reasoning-effort levels. The company highlighted massive upgrades to its Agent capabilities, surpassing V4-Pro-Preview, with native support for the Responses API format and full adaptation for Codex. Additionally, the vLLM team confirmed that serving configs from previous previews carry over, and the DSpark draft module is built directly into the weights, allowing speculative decoding to be enabled with a single flag.
Related event: DeepSeek Releases V4-Flash: Open-Source Model with Massive Agent Upgrades(80 posts)→
More from coding & agent
- EvoSkill v1.3.0 Runs Self-Improving Coding Agents Without Claude API — 0xsachi · 2026-08-01
- GitHub Launches Stacked PRs, Detailing Engineering Challenges Behind the Build — mariorod1 · 2026-08-01
- Ditching PBIX in Power BI: A Guide to PBIP and AI Agent Workflows — adnan_hashmi · 2026-08-01
- PromptLayer Launches Tool Response Mocking for Agent Testing Without Backend — Jonpon101 · 2026-08-01
- AdaMAST: Automating Agent Failure Taxonomies Boosts SWE-bench to 70.7% — berkeley_ai · 2026-08-01
- Databricks Free Edition Adds Agent Bricks and Serverless GPUs — usamawahabkhan · 2026-08-01