NetEase Youdao open-sources the LobsterAI desktop agent
NetEase Youdao open-sourced LobsterAI, packaging multi-model access, skills, scheduled tasks, and local desktop interaction into a graphical personal agent.
SEARCH
Search English reports
Browse published English reports with original sources and content provenance.
AI-generated
MiniMax releases MiniMax M2.5
An iterative upgrade to M2 that reworks the mixture-of-experts routing scheduler and improves coherence over very long contexts.
Z.ai releases GLM-5
Z.ai's fifth-generation foundation model, built on a new architecture with broad gains across general benchmarks.
OpenBMB releases MiniCPM-o 4.5
An efficiency-focused, fully multimodal open model tuned for on-device chips and offline inference.
Kuaishou releases Kling AI 3.0
Kling AI 3.0 unifies video understanding, generation, and editing, with full multimodal I/O, native audio, and stronger narrative control.
Anthropic releases Claude Opus 4.6
A late-cycle refinement of Opus, with targeted fine-tuning for a set of long-tail specialized tasks.
OpenAI releases GPT-5.3-Codex
A closed, developer-focused build that combines GPT-5's general capabilities with OpenAI's advanced code-generation stack.
Mistral releases Voxtral Mini Transcribe Realtime
Mistral releases an open-weight, low-latency streaming-transcription model.
Mistral releases Voxtral Mini Transcribe 2
Mistral releases Voxtral Mini Transcribe 2 for batch transcription, speaker diarization and word-level timestamps.
StepFun releases Step-3.5-Flash
A hundred-billion-scale open model from StepFun tuned for lower generation latency and faster token throughput.
Google opens Project Genie to AI Ultra users
Google opened Project Genie, an interactive world model, as an experimental prototype for Google AI Ultra users in the U.S.
Clawdbot rebrands to OpenClaw
OpenClaw's stable release completes the npm-package and CLI rename, entering the agent ecosystem as a self-hostable personal AI assistant that connects to chat apps and runs real tasks.
Qwen open-sources Qwen3-ASR
Qwen releases an open-weight ASR family supporting 52 languages and dialects with streaming and offline inference.
xAI launches the Grok Imagine Image API
xAI launched a revamped generation of Grok Imagine image models as a programmable API for standard image generation and editing.
Arcee AI releases 400B MoE model Trinity Large
Arcee AI releases Trinity Large, a 400-billion-parameter mixture-of-experts open model built with model-merging techniques.
Alibaba Cloud releases Qwen3-Max-Thinking
An advanced closed reasoning variant offered internally by Alibaba, with stronger deep reasoning and complex logical problem-solving.
Moonshot AI releases Kimi K2.5
Kimi K2.5 is a natively multimodal agent model continually pretrained from Kimi K2, with visual understanding, coding, and Agent Swarm support.
Google adds free full-length SAT practice tests to Gemini
Gemini adds full-length SAT practice tests backed by education-provider content, with scoring, explanations and personalized feedback.
Sarvam AI releases 2B model for India's language ecosystem
Sarvam AI releases Sarvam-2B, a compact foundation model trained for India's complex and fragmented multilingual environment.
Anthropic launches the Claude Cowork research preview
Anthropic extends Claude Code's execution model to non-coding work, letting Claude plan tasks and produce professional deliverables inside authorized folders.