Cur8 — Wednesday, August 26, 2026

Qwen3.8-Flash-Next achieves 90% efficiency of Qwen3.8 with only 50% compute cost via novel architecture.

Ox Alpha's weights released; new GLM-series model from Z.ai confirmed.

Finetuned multi-vector model outperforms general-purpose retrievers on medical data, trained in 14.5 hours on a single RTX 3090; -unsupervised checkpoints adapt better to new domains than finished siblings.

Jalapeño outperforms Nvidia Blackwell and Rubin on perf/W, achieving over 700 tokens/sec/user at concurrency 1 with single-token prediction; B0 stepping offers 25% improvement.

OpenAI CTO details how progress in hardware, algorithms, and software drives scalable AI at reduced costs.

loveholidays leverages OpenAI Codex for widespread software development, accelerating product creation.

Admin plugin for ChatGPT Work and Codex enables workspace management, including usage analysis, member/permission control, limit adjustments, and request handling.

$5M grant program funds independent research on AI's impact on user wellbeing, with open-source evaluations and benchmarks due by September 21.

AWS acquires DuckLabs; DuckDB remains open-source under MIT license. AWS aims to expand DuckDB's reach, with over 1 million daily downloads.

MCP Apps in Amazon OpenSearch Service enable inline verification of AI agent findings, reducing tab-switching and context-carrying time. The feature provides interactive visualizations alongside text responses for observability tasks like trace investigation and service topology mapping.

AI-assisted reporting cuts weekly report preparation from hours to minutes using Amazon Quick Desktop and FSx for NetApp ONTAP; governed workflow ensures trusted source metrics.

Apple and OpenAI challenge Nvidia with new hardware announcements.

DeepSeek's open-source AI harness enables self-rewriting programs, allowing customization of UI, agents, and functionalities on-demand, with efficient, reversible changes. Hundreds of plugins already created post-release.