Cur8 — Wednesday, August 26, 2026

Qwen3.8-Flash-Next achieves 3.8x higher throughput and 2.5x lower latency than previous models, cutting inference costs by 60% while maintaining accuracy. Released in October 2024.

Z.ai confirms Ox Alpha as a new GLM-series model with open weights, enhancing accessibility for research and development in large language models.

Multi-vector models outperform general retrievers on medical data. Training on 4.4M medical question-passage pairs, the model runs on a single RTX 3090 in 14.5 hours.

OpenAI's Jalapeño inference chip outperforms Nvidia Blackwell and Rubin on perf/W, achieving 700+ tok/s/user at concurrency 1 on DeepSeek R1, using HBM4 and single-token prediction without speculative decoding.

Advances in chips, compute, models, and products compound to deliver more useful intelligence at greater scale and lower cost, as explained by OpenAI CFO Sarah Friar.

Loveholidays uses OpenAI Codex to democratize software development, enabling non-experts to build tools and features quickly, accelerating product innovation across the company.

Admin plugin for ChatGPT Work and Codex enables workspace analysis, member management, permission adjustments, and handling of admin requests.

Anthropic launches $5M grant program to fund independent research on AI's impact on user wellbeing, aiming to develop open-source evaluation benchmarks. Applications due September 21, with full proposals notified by October 5.

AWS acquires DuckLabs, effective early September; DuckDB and other projects remain open source under MIT license. AWS will support DuckDB's growth, aiming to expand its impact through enhanced infrastructure and broader adoption.

AWS OpenSearch Service MCP Apps enable agentic observability by embedding interactive visualizations (trace waterfall, service map) directly in AI chat threads, eliminating tool-switching. Verification happens in the same thread with deterministic, dashboard-matching results.

Amazon Quick Desktop cuts weekly report prep from hours to minutes using AI and governed access to files on FSx for NetApp ONTAP. The solution integrates with S3, Slack, and IAM for secure, repeatable reporting workflows.

Apple and OpenAI both announced hardware updates challenging Nvidia's dominance in AI computing.

DeepSeek's new AI system allows self-modification and customization, enabling users to create agents for tasks like code review or research analysis. The system is efficient, reversible, and runs locally with no tracking.