The Tencent Hunyuan team has launched the next-generation MoE flagship model 「Hy4 preview」, featuring a total of 770B parameters with 49B activated per token. The model adopts a 78-layer backbone, 77 of which are MoE layers, each containing 256 routed experts and one shared expert. It supports a context length of up to 1 million tokens, targeting use cases in software engineering, office analytics, game development, and scientific research. Project details and architectural documentation have been published on the official GitHub page, where Tencent claims improvements over Hy3 in reasoning and vision-language tasks. Developers can review specific design and usage information.

OpenAI Developers has introduced 「Appshots」, providing ChatGPT Work and Codex with visual and textual context from users' current application interfaces. This capability allows models to understand on-screen content and act accordingly, aiming to eliminate the need for manual copying of interface data. It can be used for analyzing chat threads, identifying app states, assisting debugging, and extending static screenshot understanding into real-world application workflows. Official materials have not yet disclosed availability scope, API access methods, pricing, or rollout timeline.

OpenAI has released the 「Rosalind Workbench」, designed for protein structure and genomic data analysis, integrating AI models, visualization tools, and sequencing pipelines into a unified platform. The system aims to bridge raw research data, model inference, and result interpretation, reducing the burden of switching between multiple bioinformatics tools. It employs domain-specific models to support tasks such as protein folding and sequence alignment. However, official documentation does not specify pricing, user eligibility, data sourcing requirements, or official release date.

Grok Bot has gained online purchasing capabilities, allowing users to link payment cards via Stripe and delegate product selection and checkout processes to the bot. Public demonstrations show the agent browsing e-commerce platforms, selecting items, and initiating payments—extending its functionality from information retrieval and web navigation into real transactions. Payment card binding and transaction records are managed by Stripe, but current materials do not clarify per-transaction limits, refund mechanisms, supported merchants, geographic availability, or liability and human confirmation rules in cases of erroneous purchases or misselection.

Replit has introduced 「Growth Skills」, integrating customer acquisition, campaign launches, payments, and user behavior tracking into Replit Agent through partnerships with analytics, marketing, and sales platforms. This solution enables developers to manage the entire workflow—from product creation to market entry—within a single environment. With partners like PostHog, users can visualize the complete customer journey from discovery and engagement to conversion and retention. The agents can also assist in lead generation and marketing optimization experiments. However, official materials have not yet disclosed the full list of partners, pricing, or availability scope.

Google has integrated Gemini into Waymo vehicles, offering passengers a hands-free in-car voice assistant capable of retrieving information and controlling cabin functions. Google emphasized that Gemini operates independently from the autonomous driving system 「Waymo Driver」, ensuring that passenger interactions do not interfere with vehicle control decisions. This integration brings general-purpose voice AI into autonomous mobility scenarios. However, details on available cities, supported vehicle models, specific cabin controls, and full rollout schedule remain undisclosed.

LangChain has announced MCP support within its open-source framework, using FastMCP as a foundational component for server and client development to improve interoperability among models, tools, and execution environments. Developers can now test early implementations in the experimental release langchain==1.4.0a2 and explore building or connecting MCP components within existing LangChain applications. The update remains in preview, primarily intended for feedback collection and compatibility testing. A stable release date and comprehensive migration guide have not yet been provided.

Cohere has launched 「Parse」, an enterprise-grade document intelligence tool that converts complex files—including text, tables, and other multimodal content—into structured data for retrieval, analysis, or downstream AI workflows. The product supports nine major languages and is priced at $1.5 per 1,000 pages processed, offering predictable cost modeling for large-scale digitization of contracts, reports, and archives. However, official materials do not disclose supported file formats, per-file size limits, data residency options, accuracy metrics, or a dedicated landing page.

Maintainers of the OCaml project reported that approximately 10 minutes after sharing a patch in a public repository, their website began receiving automated probes targeting percent-encoded path traversal vulnerabilities—indicating active monitoring bots scanning public code changes. Additional tests suggest modern AI agents can rapidly generate functional exploit methods from vague vulnerability hints, significantly narrowing the window between patch disclosure and malicious exploitation. This incident highlights growing pressure on traditional open-source disclosure, delayed release, and patch coordination practices, though no actual breach or data leak was confirmed.

Google DeepMind has released the open model 「DiffusionGemma」, which refines 256-token blocks in parallel through denoising, differing from traditional autoregressive, token-by-token generation in large models. According to reports, it is based on the Gemma 4 26B architecture and achieves generation speeds exceeding 1,000 tokens per second on a single H100 GPU—up to four times faster than comparable Gemma models. It supports a 256K context and over 140 languages. Licensed under Apache 2.0, the model is suitable for commercial and local experimentation, though performance figures originate from secondary technical articles.