ModelBest says its open MiniCPM family has surpassed 50 million downloads as the Chinese lab prioritizes efficient, on-device AI.
⚡ Quick hits
AgentsHuawei’s Xiaoyi ranked first overall among smartphone agents in China Telecom Research Institute’s 2026 AITMark evaluation.IT之家 ↗
ModelsZhipu confirmed that the anonymous Ox Alpha model is a new GLM-series version and said it will release the weights.IT之家 ↗
AIGCAlibaba’s Wan 3.0 became available on ArtArch AI with limited-time free generations and support for cinematic motion and scene consistency.X @Alibaba_Wan ↗
AgentsManus restored normal service and opened data restoration for users who had previously created backups.X @ManusAI ↗
SafetyTrustShiftProbe introduced a benchmark and defenses for staged trust attacks in which malicious MCP servers build credibility before changing behavior.arXiv ↗
ResearchResearchers found that agentic scaffolding can amplify large language models’ tendency to agree with users’ incorrect or biased premises.arXiv ↗
ResearchA new reinforcement-learning method trains generalist robot policies to use inference delays productively instead of remaining idle while awaiting actions.arXiv ↗
ResearchESQ-Bench evaluates whether text-to-SQL systems generalize across enterprise database dialects while detecting syntactically valid queries with silently divergent meanings.arXiv ↗
FundingIndian voice-agent startup Ringg secured a $10 million Series A extension from Peak XV after reaching 20 million monthly call attempts.TechCrunch ↗
Open SourceOllama v0.33 added one-toggle Claude Desktop configuration, enabling Ollama to serve as a gateway for local and cloud models with built-in web search.X @ollama ↗
AgentsArtificial Analysis now assigns zero credit to reward-hacked trials in version 1.4 of its Coding Agent Index.X @ArtificialAnlys ↗
IndustryOpenAI data-center chief Chris Malone departed after an infrastructure reorganization changed the reporting structure for the company’s buildout team.TechCrunch ↗
AIGCArtificial Analysis measured Breeze TTS 2 at 45 characters per second, versus 102 for open-weight speech speed leader Fish Audio S2 Pro.X @ArtificialAnlys ↗
FundingKeenable emerged from stealth with $26 million to build a web index and low-latency search APIs optimized for AI agents.SiliconANGLE ↗
AgentsOpenAI released an Admin plugin enabling ChatGPT Work and Codex administrators to manage users and permissions conversationally.IT之家 ↗
IndustryFigma increased AI add-on credits across every plan, doubling Pro allowances and raising Organization and Enterprise allocations by 60%.Figma ↗
IndustryNvidia previewed Dynamo shadow engine recovery, which maintains a standby LLM engine to reduce lost inference capacity after crashes.NVIDIA AI ↗
AIGCAdobe added Animation Generator to Firefly, turning children’s ideas and other prompts into animated visual creations.Adobe Firefly ↗
ModelsTogether AI added Qwen3.8-27B support for fine-tuning and deployment through dedicated model inference infrastructure.Together AI ↗
ResearchGoogle introduced AgentHands, an XR research prototype generating synchronized, spatially grounded hand gestures for conversational agents.Google Research ↗
Portable Computer keeps agent planning, file analysis and tool execution on a DGX Spark, escalating to cloud models only with permission.
⚡ Quick hits
AgentsGemini’s macOS app added voice-driven dictation, file summarization, and text rewriting directly inside other application windows.Gemini ↗
ResearchGitHub Next proposed a “knowledge compressor” for rigorously shrinking accumulated specifications and reducing the context consumed by coding agents.GitHub Next ↗
OpenAI says its first inference chip delivers lower latency and more work per watt across three large open-weight models.
⚡ Quick hits
AgentsAnthropic unified Claude’s memory across chat and Cowork, while letting users inspect, edit, delete, or disable remembered topics.Claude ↗
AIGCRecraft redesigned Studio around Create, Edit, and Chat tabs while retaining vector editing, styles, and mockup workflows.Recraft ↗
AIGCHeyGen removed LiveAvatar API concurrency limits and advertised scalable full-body 1080p avatar sessions starting at one cent per minute.HeyGen ↗
Open SourceIBM released Apache-2.0-licensed Granite 4.2 reasoning models in 3B, 8B, and 30B sizes, featuring tool calling and up to 512K context.Hugging Face ↗