Cur8 — Thursday, August 27, 2026

Qwen3.8-Flash-Next: 125B token, 6B active multimodal MoE model; 72.5GB and 78.9GB quantized versions available.

Open-source AI executive assistant Open Executive launched, offering eight specialist agents with episodic memory and scheduling capabilities; uses Anthropic Claude models.

OpenAI outlines measures to bolster AI model security post-Hugging Face breach; focuses on enhanced monitoring and alignment.

Brazil gains expanded access to OpenAI's resources for AI development and adoption.

Expanding to 55 U.S. school systems, ChatGPT for Teachers now supports over 100,000 educators with secure AI tools and training.

ChatGPT aids continuous learning; OpenAI report highlights student-educator usage extending beyond classrooms.

Nvidia may acquire Hugging Face for $13B, boosting its foothold in open-source AI ecosystem. Deal could drive more workloads onto Nvidia chips but may complicate Hugging Face's neutrality.

Double-blind AI evaluations launched by DeepMind using cryptographic "box" to prevent benchmark contamination; Gemini Flash Lite tested with Singapore AI Safety Institute, OpenMined, AVERI, MLCommons.

Gemini 3.5 Transcribe achieves WER of 4.0% (streaming) & 2.6% (non-streaming), supports 85+ languages, multi-speaker ID; available via Gemini API, Google AI Studio, & Enterprise platforms.

Security breach at Hugging Face exposed 1% of private datasets; company enhancing security measures.

$399 open-source biped robot, Microduck, trains via reinforcement learning; pre-orders open for Q4 2026 delivery.

Amazon Bedrock AgentCore Evaluations standardizes agent evaluation across frameworks using OpenTelemetry, supporting Strands Agents, LangGraph, OpenAI Agents SDK, LlamaIndex, Google ADK, and Claude Agent SDK. It requires session.id grouping and message content inclusion for accurate scoring.

Advanced data strategies for supervised fine-tuning: Evaluate dataset readiness with learning curve analysis (2,000 samples as a starting point), optimize data subset selection, augment data through reasoning traces or synthetic demonstrations, and mix data to prevent catastrophic forgetting. Data mixing ratios vary by task and model size; monitor token-level ratios.

- Evoke: Real-time interactive world generation with 14B parameters. - 4D Anyone: 3D character reconstruction from single videos, 12GB model size. - Sense Nova U 1.58B: Open-source image generator/editor, native 4K resolution, 50GB model. - Bernini v2: Omnimodal video editor, 180GB model size. - Ornith 1.5: Self-improving AI models up to 397B parameters, outperforming larger models. - Audio8TS: Tiny text-to-speech generator with 0.1B parameters, 1.7GB total size.