AI Updates on 2026-02-25

Alibaba releases Qwen 3.5 Medium Series including 35B-A3B, 122B-A10B, and 27B models with 1M+ context support on consumer GPUs @Alibaba_Qwen
Qwen3.5-35B-A3B surpasses previous Qwen3-235B models at 6x smaller size, demonstrating architecture and data quality improvements over parameter scaling @Alibaba_Qwen
Inception Labs launches Mercury 2, first reasoning diffusion LLM delivering 5x faster performance than speed-optimized autoregressive models @StefanoErmon
xAI debuts Grok-4.20-Beta1 ranking #1 on Search Arena and #4 overall in Text Arena, scoring 1492 and closing gap to Gemini 3.1 Pro @arena
Anthropic acquires Vercept to advance Claude's computer use capabilities for autonomous task execution @AnthropicAI
Claude Code launches Remote Control feature allowing users to start terminal sessions and continue from mobile devices @claudeai

Perplexity launches Computer platform orchestrating 19 models for end-to-end project management with usage-based pricing for Max subscribers @perplexity_ai
OpenRouter processes 13 trillion tokens in week ending February 9, up from 6.4T in early January with 4 of top 5 models being open weight @ai
MatX raises $500M Series B led by Jane Street to develop LLM chip combining low latency of SRAM-first designs with long-context HBM support @reinerpope
a16z leads Quiver AI's $8.3M seed round for frontier SVG generation model producing production-ready vector graphics from images and text @joanrod_ai
Chariot Defense raises $34M Series A led by a16z to scale battlefield power systems with Amphora products supporting U.S. Army units @chariotdefense
Sierra partners with Next fashion retailer going live in 83 countries within 6 weeks using AI-powered customer service agents @btaylor

Google launches Gemini automations on Android allowing AI to navigate apps and complete multi-step tasks like grocery ordering autonomously @GeminiApp
Waymo announces testing expansion to Chicago and Charlotte as autonomous vehicle deployment accelerates across U.S. cities @TechCrunch
Google DeepMind's Project Genie generates navigable environments from single prompts to create safe testing grounds for training AI agents @GoogleDeepMind
Notion enables custom agents to run arbitrary TypeScript code allowing connections to any external service or API @_clem

Google's Aletheia agent powered by Gemini 3 Deep Think autonomously solves 6 of 10 FirstProof challenge problems, best result on mathematician-level tasks @quocleix
Princeton research finds two years of AI capability progress produced only modest reliability gains across 12 dimensions including consistency and robustness @random_walker
Snap Research introduces RefVFX for tuning-free visual effect transfer between videos using finetuned Wan 2.1 model with compositional inference @rsalakhu