Cur8 — Wednesday, August 26, 2026

Alibaba’s Qwen3.8-Flash architecture achieves comparable performance to Llama 2 70B with just 7B parameters, drastically reducing inference costs and memory footprint—demonstrating a significant leap in cost-efficiency for large language models. Initial benchmarks show promising results across various tasks.

Z.ai’s Ox Alpha, a new 180B GLM-series model, rivals Llama 2 70B on benchmarks; they'll release open weights starting November 23rd. This accelerates accessible large language model development and research.

**Hugging Face's Sentence Transformers enables multi-vector embedding models to outperform existing retrievers.** A new blog post details training and finetuning these models, achieving a +0.0311 NDCG@10 improvement on medical data using 25k pairs & a single RTX 3090; pre-supervised checkpoints adapt best for domain specialization.

OpenAI’s new Jalapeño inference chip surpasses Nvidia Blackwell and AMD Rubin in performance-per-watt, achieving industry-leading token throughput across various models—even without key optimizations. Built in just 16 months with Broadcom, it utilizes HBM4 and demonstrates AI accelerating chip design, signaling a shift towards power efficiency as the primary constraint.

Lower costs & increased scale: OpenAI's advancements in hardware, models, and infrastructure are compounding, enabling broader access to AI capabilities. This holistic approach optimizes performance across the entire AI stack.

Loveholidays empowers non-technical employees to build applications using OpenAI Codex, accelerating product creation and democratizing software development within their organization. This expands development capacity beyond traditional engineering teams.

ChatGPT Work & Codex admins now centrally manage workspaces via a new plugin, streamlining user access, permissions, and resource allocation. This boosts operational efficiency and control for enterprise deployments.

Anthropic launches a $5 million grant program for independent research evaluating AI's impact on user wellbeing, providing funding and access to models. Evaluations must be rigorous, expert-validated, and reflect real-world conversational contexts by September 21.

AWS acquired DuckLabs in September for an undisclosed sum, ensuring DuckDB and related open-source projects (DuckLake, Quack) remain free under MIT license. AWS aims to expand DuckDB’s reach and integrate it into new data services, leveraging its infrastructure and scale.

AWS OpenSearch Service now supports MCP Apps, delivering interactive observability visualizations (trace waterfalls, service maps) directly within AI agent chat threads. This eliminates tab-switching for verification, streamlining investigations and boosting efficiency by ~80%—a key bottleneck for local agentic setups.

**AWS Quick Desktop & FSx for NetApp ONTAP slash weekly report prep time.** AI-assisted reporting now integrates governed files on FSx for ONTAP, reducing hours of manual effort to minutes with a new desktop application and Slack integration. Controlled access and human review ensure data trust.

Apple’s updated Mini/Studio Macs now feature Apple Silicon with improved Neural Engine performance, challenging Nvidia's GPU dominance. OpenAI released "Jalapeño," a faster, cheaper language model, also pressuring Nvidia's AI compute market share.

DeepSeek released a free, open-source AI harness enabling self-extending functionality: users prompt new tools directly, which the system creates and integrates—no coding required. The architecture includes reversible changes via "tickets" for automated cleanup, fostering rapid customization and efficiency.