Daily AI News | Mistral opens a preview of Large 4
# Daily AI News | Mistral opens a preview of Large 4
## Lead Story
### Mistral opens a preview of Large 4
Mistral Large 4, also called Le Chonk, is being launched in preview for developers, cybersecurity leaders, and government authorities. Media reports describe it as a one-trillion-parameter model; performance comparisons are attributed to Mistral.
**AI take:** The current preview is limited; self-hosting depends on the planned weight release.
## More News
### Claude launches a Google Workspace sidebar add-on
Anthropic has begun a public beta of Claude for Google Workspace, letting users open Claude beside Docs, Sheets, and Slides to read and edit files. A separate connector can create or modify Workspace files from Claude chats; both features are for Pro, Max, Team, and Enterprise plans.
**AI take:** It brings Claude into existing office workflows, but access is limited to paid-plan users during beta.
### Google releases five-modality EmbeddingGemma 2
Google DeepMind released EmbeddingGemma 2, an open multimodal embedding model that maps five modalities into a shared representation space. Reports describe it as a lightweight option for retrieval and on-device development; it is not a general-purpose chat model.
**AI take:** It is aimed at cross-modal search workflows rather than conversational use.
### Hark Pro personal assistant enters wide release
Hark Pro has entered wide release with free access and a paid tier for heavier use. The assistant can navigate the web and connect with tools such as email, calendars, and files; its company also describes a longer-term vision for AI computing devices.
**AI take:** Its value will depend on completing cross-app tasks reliably, not just showing its work.
### Infineon completes its acquisition of C2i Semiconductors
Infineon said it completed its acquisition of India's C2i Semiconductors. C2i develops software-defined multiphase controllers and smart power stages for AI data centers; Infineon says the deal adds those products and expertise to its power business.
**AI take:** More AI compute also requires power-delivery hardware, not just accelerator chips.
### KT launches a product-information plugin in ChatGPT
KT has launched an official product-information plugin in ChatGPT. Users can ask natural-language questions to find and compare the carrier's mobile, broadband, and IPTV plans instead of searching its website or app first.
**AI take:** It adds a discovery channel for Korean customers; it does not change the plans' prices.
### llama.cpp 0.6.0 adds a new API and Apple GPU acceleration
ggml-org released llama.cpp 0.6.0 with a new batch API, support for additional models, Apple GPU acceleration, and web-interface updates. The runtime is available free from GitHub under the MIT license.
**AI take:** Runtime improvements can make it easier for developers to test newer models locally.
### Minerva shows a remotely operated humanoid for hazardous work
Minerva Humanoids has built Roger, a humanoid intended for explosive ordnance disposal and hazardous-energy work. An operator controls it remotely through a VR headset while onboard software handles balance, navigation, and recovery; the report describes a prototype, not a deployed service.
**AI take:** The design keeps specialists in control while moving them away from the immediate hazard.
### Precisely opens a preview of its AI-ready data platform
Precisely opened a public preview of a unified data-management platform spanning integration, quality, governance, location intelligence, master data, and customer communications. General availability is planned for early 2027, and the company says the platform can work across existing environments without a prior migration.
**AI take:** Teams can assess cross-system data management now, but the full launch is still planned for next year.
### Qatar's Qai and MEEZA agree on a GPU infrastructure deployment
Qatar's AI company Qai plans to deploy GPU infrastructure with data-center provider MEEZA at a certified facility in Qatar. MEEZA will provide operations, monitoring, and lifecycle services; the report does not disclose capacity or specific workloads.
**AI take:** The partnership adds a regional compute project, though its scale and activation date remain undisclosed.
### RoboParty unveils RP1 and plans to open its humanoid stack
RoboParty unveiled RP1, a humanoid platform for researchers, educators, and developers, and plans to progressively open its mechanical designs, actuators, motion control, simulation, SDK, and training resources. The materials are planned for staged release, not all immediately available.
**AI take:** If delivered, the open materials could lower the hardware barrier for embodied-AI research.
### Sierra and partners announce a plan for an open personal-agent protocol
Sierra and partners including Meta, Genesys, and Shopify announced the Personal Agent Protocol, intended to define interactions between personal agents and businesses, including authorization and business-side visibility. The v0.1 specification is planned for late October and is not yet available.
**AI take:** If adopted, the standard could make user authorization central to how agents connect to services.
### Underdog opens an invite-only beta for its on-device assistant
Founder Sigil Wen opened an invite-only beta of Underdog, a personal assistant that runs wholly on-device and connects with a user's digital tools. Access remains limited, and its privacy advantage is a product claim rather than an independently measured comparison.
**AI take:** Local execution can reduce the need to send data off-device, but access is still narrow.
### UK Government accepts 44 healthcare-AI recommendations
The UK Government has accepted all 44 recommendations from the National Commission into the Regulation of AI in Healthcare. The commission's report appeared on September 10; acceptance is a government response, not the enactment of new regulations.
**AI take:** Providers and developers gain a policy signal, but not a new compliance rule yet.
### Exa Enterprise AI plans Buddy Agent for late October
Exa Enterprise AI announced Buddy Agent for its Exabase AI service, with availability planned for late October. The feature is intended to plan and execute work steps, and a separate organizational-knowledge feature based on internal documents is also planned.
**AI take:** If delivered as described, it could reduce the need for employees to turn work into detailed prompts.
### Doosan wins two physical-AI projects worth about KRW 98.9 billion
Two Doosan Robotics projects were selected for government R&D support: an on-device-AI collaborative robot and an AI welding system for nuclear-equipment manufacturing. Both are still in development.
**AI take:** The projects bring edge AI into industrial robotics, but deployment depends on the R&D still ahead.
### Ray-Ban Meta Gen 3 glasses set for an October 12 Korea launch
Ray-Ban Meta Gen 3 is scheduled to launch in Korea on October 12, starting at KRW 740,000. The glasses combine a 3K camera with Meta AI, and the company lists battery life of up to nine hours.
**AI take:** The update focuses on camera and battery; real-world use will show how much those changes matter.
### Infor expands its industry AI platform with more agents
Infor announced an expanded Industry AI architecture and a next-generation Velocity Suite, adding adaptive experiences, a customer-wide governance model, and more industry agents. The cited finding that 68% of respondents see gaps in generic AI comes from Infor's own survey.
**AI take:** Enterprise software is adapting general AI to industry workflows; governance and deployment results still need to be tested.
### Anthropic offers eligible startups a free year of Claude Team
Anthropic expanded Claude for Startups with a free year of Claude Team for qualifying companies, covering up to five premium seats, plus $1,000 in API credits. The offer is eligibility-limited rather than a general price cut.
**AI take:** For early teams, the seats and credits can lower the cost of testing collaborative and API workflows.
### Musubi releases open-weight PolicyLM-1.7B for content moderation
Musubi announced PolicyLM-1.7B, an open-weight decision model for real-time content moderation. The report says it is designed to apply plain-English policies to messages in under 50 milliseconds.
**AI take:** Open weights let platforms test it themselves; real latency and moderation quality remain to be measured.
### Incognia launches AI-agent risk detection for financial firms
Incognia launched AI Agent Detection for financial institutions to identify agent activity and assess whether individual actions are authorized by the person behind them. The company also announced expanded web-risk capabilities; no independent performance data was provided.
**AI take:** As finance becomes more automated, identifying an agent and authorizing each action are separate checks.
## Rumors
