This page may contain stale information. Last updated: 2026-08-14
Overview
Trend of open-weight / low-cost models competing on agentic and coding benchmarks via post-training rather than architecture leaps — exemplified by deepseek-v4-flash 0731.
Timeline
-
2026-08-14: Z.ai delays glm-5-3 weights citing emergent exploit-chain capability (2026-08-14-zai-glm-53-the-decoder)
-
2026-04-24: V4 Flash preview
-
2026-07-31: Official Flash-0731 graduation (2026-08-01-deepseek-v4-flash-0731-official-release)
Key Players
Analysis
Vendor harness scores must be caveated; Responses API / Codex ergonomics may matter as much as raw bench numbers for developers.