EASYHUB JOURNAL / 2026-06
AI news in June 2026
14 stories, ordered by their original dates. Open a title for its summary, image and source link.
- Anthropic
Introducing Claude Sonnet 5
Anthropic introduces Claude Sonnet 5 with stronger planning, browser and terminal use, coding and professional workflows. The announcement also discusses safeguards and API access; current pricing and account limits should be checked separately.
- Google
Start building with Nano Banana 2 Lite and Gemini Omni Flash
Google brings the lighter Nano Banana 2 Lite image model and Gemini Omni Flash video generation and conversational editing to developers. They connect image iteration with video creation, with feature and regional limitations documented separately.
- Ollama
Faster Gemma 4 on MLX with multi-token prediction
Ollama 0.31 uses MLX and multi-token prediction to accelerate Gemma 4 on Apple Silicon, with a draft model proposing tokens for verification. Reported speedups come from a particular coding benchmark and should not be generalized to all devices and workloads.
- IT之家
DSpark optimizes speculative decoding for concurrent serving
DSpark shares draft-model training and evaluation tools for balancing latency and throughput under concurrent serving. Reported improvements depend on the production baseline and workload, not every GPU configuration.
- IT之家
JoyAI-VL-Interaction opens a continuous video-interaction stack
JD combines video perception, voice and tool interfaces in a modular interaction stack. Suggested care-related applications are possible directions, not evidence that the model provides reliable supervision.
- IT之家
Miloco 2.0 brings persistent memory and task orchestration to smart homes
Miloco combines perception, memory and tasks for connected homes. Cameras and household-member information raise privacy concerns, so deployment requires clear consent, access limits and retention choices.
- IT之家
QoderWork adds editable local memory, reflection and skill suggestions
QoderWork organizes conversations into short- and long-term memory, removes redundant information and suggests skills from repeated tasks. Memory is stored in editable local Markdown, with snapshots before reflection and user approval before adopting skill suggestions. The branded “consciousness” feature refers to memory management, not evidence of sentience.
- IT之家
Yuanbao adds ima knowledge sources with links back to original material
Yuanbao adds ima knowledge content as an answer source. Citation cards open the original material and its library, which users can join for further questions. The report also describes links with WorkBuddy and Tencent Docs. It does not establish automatic access to every user’s private library.
- IT之家
MiMo Code introduces an OpenCode-based terminal coding assistant
MiMo Code adds localized setup, model connections and task status to an OpenCode-derived terminal assistant. Open-source client licensing and model-service quotas are separate, with free promotions limited to their stated period.
- IT之家
MusaCoder targets PyTorch-to-CUDA and MUSA kernel generation
MusaCoder opens specialized models for kernel generation and documents its training platform. Generated kernels still need numerical and real-hardware performance validation; successful compilation alone is insufficient.
- IT之家
WorkBuddy Enterprise links agents with documents, storage and team knowledge
Tencent introduces WorkBuddy Enterprise and Agent Suite with expert agents, project collaboration and administrative controls. A unified account connects Tencent Docs, storage and Lexiang for shared editing and retained knowledge. The same event updates ADP 4.0 and ClawPro for agent building, distribution and centralized management.
- IT之家
Kimi Work Beta brings local file and browser tasks to Mac and Windows
Kimi Work Beta uses Kimi Code underneath a desktop interface for files, browsers, skills and scheduled tasks, delivering documents, spreadsheets and presentations. It inherits specialist online-agent capabilities and supports parallel task execution. This announcement is an internal beta in Mac and Windows test clients, not universal access.
- IT之家
PaddleOCR-VL-1.6 keeps a compact architecture for complex documents
PaddleOCR-VL-1.6 retains a 0.9B architecture while improving tables, formulas, rare characters and difficult photographed or curved documents. The announcement’s 96.33 figure is tied to a specific OmniDocBench evaluation, not a universal accuracy guarantee. Representative business documents still require validation.
- IT之家
Coze 3.0 brings local and cloud agents into shared team projects
Coze 3.0 updates mobile, desktop and web clients with agent teams and project-based task allocation. Users can connect local Claude Code, Codex CLI or OpenClaw agents, or create agents running on cloud computers, bringing different execution environments into a shared project instead of a single chatbot conversation.