EASYHUB JOURNAL / 2026-07
AI news in July 2026
26 stories, ordered by their original dates. Open a title for its summary, image and source link.
- ByteDance Seed
ByteDance Seedance 2.5 expands narrative video and reference editing
ByteDance Seed introduces Seedance 2.5 for longer narrative video, richer multimodal references and targeted editing. The announcement describes a rollout to Dreamina and Doubao Pro, while the Volcengine API is still forthcoming at publication.
- DeepSeek
DeepSeek V4 Flash API enters public testing
The July 31 changelog describes a post-training update to the V4 Flash API and compatibility with the Responses API and Codex. At that time the change did not update Pro or the app and web models. Later replacements are separate events.
- MiniMax
MiniMax H3 brings unified multimodal video generation
MiniMax introduces H3, a generative system using text, images, video and audio as a shared context. It supports video generation with native stereo audio and multimodal editing. This entry covers the launch; the subsequent release of weights is recorded separately.
- IT之家
WorkBuddy adds shared AI editing for online and local documents
WorkBuddy can open Tencent Docs and local Word, Excel, PowerPoint or Markdown files in its document editor. Users select text, tables or slide elements and ask AI to edit the original content, then continue editing or commenting. Online documents also support collaboration with colleagues.
- LightOn
mDenseOn with the mLateOn: Open Multilingual, Long-Context, and Code Retrieval Models
LightOn releases mDenseOn and mLateOn retrieval models with work on multilingual, long-document and code search, alongside models, data and training code. Reported comparisons are author evaluations and should be validated on the intended collection.
- OpenAI
How GPT-5.6 fuses frontier intelligence with frontier efficiency
OpenAI examines efficiency work across training, inference infrastructure and agent harnesses for GPT-5.6. The post discusses reducing processing and coordination costs and is a technical explainer rather than a separate model launch.
- Moonshot AI / Kimi
Kimi K3 releases weights, its report and infrastructure tools
Kimi's open day publishes K3 model weights and its technical report, together with MoonEP, FlashKDA and AgentEnv infrastructure. Model and code repositories have their own deployment and licensing terms; an open code release does not establish a single license for every asset.
- IT之家
WorkBuddy desktop arrives in the HarmonyOS computer app store
WorkBuddy arrives in App Gallery on HarmonyOS computers for documents, coding and design. The listing identifies native HarmonyOS support. Alongside the previously launched HarmonyOS mobile app, users can initiate computer tasks from a phone, extending the workspace across more devices.
- IT之家
CodeBuddy NPC executes cloud development tasks around issues, pull requests and CI
CodeBuddy NPC uses Tencent CNB repositories, issues, pull requests and CI results to plan work, edit code, submit changes and iterate on tests. Organizations can configure models or assemble NPC teams. The system reduces model-call overhead; it does not literally consume zero tokens.
- OpenAI
Introducing OpenAI Presence
OpenAI Presence combines model reasoning with enterprise knowledge, permissions, approvals and escalation to people for customer and internal workflows. It is introduced through a limited enterprise deployment program, not as a universally self-service assistant.
- ByteDance Seed
Seed Audio 1.0 brings speech, sound effects and ambience into one scene
ByteDance Seed introduces Seed Audio 1.0 to generate dialogue, effects and ambience as a coordinated sound scene. It supports timing controls, reference voices and multilingual output. The Chinese announcement points to a Volcengine experience entry rather than promising every future capability is already available.
- IT之家
WorkBuddy mobile apps support cloud tasks and connected computers
WorkBuddy launches mobile apps for Android, iOS and HarmonyOS. Users can run cloud tasks with skills, experts, schedules and project management, or connect a computer for desktop work. Inputs include text, speech, photos and files, with imports from Tencent Docs and ima.
- Moonshot AI / Kimi
Kimi K3 launches for long-horizon coding and multimodal work
Moonshot introduces Kimi K3 for extended coding, reasoning and knowledge work, with native vision and a million-token context. Access spans Kimi products and its API. The announcement separately schedules the full weight release, rather than claiming downloadable weights on launch day.
- IT之家
Lingxi Professional organizes AI office work around project context
Lingxi Professional combines project context with WPS document operations, including native formulas, charts and document interfaces, to produce editable collaborative outputs. It also connects external tools, browser actions and local commands. Those execution features depend on the relevant environment and permissions, rather than guaranteeing unattended completion of every task.
- Thinking Machines Lab
Inkling: Our Open-Weights Model
Thinking Machines releases the multimodal open-weights Inkling model with customization and fine-tuning through Tinker. The announcement emphasizes controllable reasoning and adaptation, while the accompanying Inkling-Small is still a preview at that point.
- OpenAI
ChatGPT is now a partner for your most ambitious work
ChatGPT Work gathers context across apps and files, breaks goals into steps and creates documents, spreadsheets, presentations or web apps. Users can follow progress, add direction and approve important actions; external access still depends on authorization.
- OpenAI
GPT-5.6 is now the preferred model in Microsoft 365 Copilot
OpenAI describes GPT-5.6 becoming the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat and Cowork. This entry concerns an office-product integration, with access governed by the relevant service rollout.
- OpenAI
GPT-5.6: Frontier intelligence that scales with your ambition
OpenAI launches the GPT-5.6 family with Sol, Terra and Luna tiers for different capability, cost and latency needs, alongside higher-effort and multi-agent workflows. This entry uses the original announcement date, not subsequent pricing-update dates.
- ByteDance Seed
Seedream 5.0 Pro moves from image generation toward controllable design
ByteDance introduces Seedream 5.0 Pro with improvements to interactive editing, image combination, text rendering and design-oriented generation. The release focuses on fitting generative images into practical creative workflows; access depends on the product rollout described by the publisher.
- IT之家
Claude Cowork tests web and mobile access with cross-device task continuity
Cowork begins web and mobile testing so users can follow desktop-started tasks, receive results and approve critical steps across devices. Cloud scheduled tasks can run while personal devices are offline. The initial rollout targets Max users; cross-device access does not imply automatic access to all local files.
- Hugging Face
Native-speed vLLM transformers modeling backend
Hugging Face describes performance work connecting Transformers model implementations with vLLM batching and optimized attention. This reduces duplicate model-porting effort, while practical speed depends on the tested architecture, hardware and serving configuration.
- Mistral AI
Robostral Navigate: single-camera AI navigation
Mistral introduces Robostral Navigate for robot navigation from ordinary RGB images and language instructions. The post discusses simulation training and generalization across embodiments, without establishing reliable autonomous behavior in every real-world environment.
- Meta AI
Introducing Muse Image and Muse Video
Meta introduces Muse Image with reference-based editing and tool-assisted generation, alongside an early look at Muse Video with audio. The image model reaches specified products, while the video model is still a preview in this announcement.
- IT之家
Fun-ASR-Realtime expands multilingual streaming recognition via API
Fun-ASR-Realtime expands to 30 languages and 16 dialects through Alibaba Cloud’s API, targeting streaming transcription and language switching in meetings or support. The report also discusses an offline variant. Its benchmark figures should not be substituted for the real-time service’s performance in every environment.
- Tencent
Tencent Hunyuan Officially Releases Hy3, Advancing Agent Capabilities and Deeper Product Integration
The formal Hy3 release improves data and reinforcement learning over the April preview, with 295B total parameters, 21B active parameters and 256K context. Open weights accompany integrations with WorkBuddy, CodeBuddy, Yuanbao and ima, plus a Tencent Cloud TokenHub API for coding, office work and multi-step tool tasks.
- Mistral AI
Leanstral 1.5: Proof Abundance for All
Mistral introduces Leanstral 1.5 for Lean 4 proof engineering, iterative feedback and code verification, with open weights and API guidance. This entry follows the July 2 announcement date rather than differing registration dates in other model documentation.