AI Weekly: Compact Models and Platform Upgrades
Compact multimodal models like Qwen3.5 Small and Phi-4 excel on-device; Claude, Gemini, GPT-5.x add memory, tasks, and 1M-token reasoning.
Compact Models Enable On-Device AI
Alibaba's Qwen3.5 Small family delivers strong multimodal performance optimized for edge devices. Microsoft's open-source Phi-4-Vision-Reasoning handles vision-language, math, and science tasks efficiently in small sizes. Google's Gemini 3.1 Flash-Lite prioritizes speed and cost-efficiency within the Gemini 3 lineup. xAI's Grok 4.20 Beta 2 improves instruction following, cuts hallucinations, and adds LaTeX for scientific text. These releases make high-capability AI viable for mobile and resource-constrained apps without cloud dependency.
Major Platforms Add Automation and Productivity Tools
Anthropic's Claude now offers free Memory import from other chatbots, Scheduled Tasks in Claude Code desktop for recurring automation, and enhanced Skill Creator for testing agent benchmarks. Google's AI Mode in Canvas supports writing/coding in search; NotebookLM generates cinematic video summaries from notes using Nano Banana Pro and Veo 3 (Gemini Ultra only); LTX Studio dubs/captions videos in 175 languages with lip sync (enterprise). OpenAI's GPT-5.4 provides 1M-token context, better coding/tool use; GPT-5.3 Instant reduces hallucinations and moralizing; ChatGPT for Excel builds models in sheets (GPT-5.4-powered); Codex app on Windows manages collaborative agents for long tasks. Perplexity's Skills teach repeatable actions in Perplexity Computer.
Emerging Research and Content Tools
Anthropic rolls out Voice Mode speech-to-text in Claude Code (5% users initially) and reports early hiring slowdowns in AI-exposed roles. OpenAI previews Codex Security for vulnerability detection/patching and studies AI's impact on learning retention/motivation. For content creators, the paid bonus introduces Article Visualizer—a free Google Opal app that analyzes text/URLs to suggest 5 visual concepts (charts, diagrams, infographics), then generates downloadable graphics like app comparisons (e.g., Obsidian vs. Notion). Pair it with Data Narrator for data viz to enhance reader comprehension without design skills.
This thin news roundup lists releases without deep analysis or build guides—scan for tools like compact models to prototype on-device features today.