Daily AI News | Meta releases its Muse Wonder model family
# Daily AI News | Meta releases its Muse Wonder model family
## Lead Story
### Meta releases its Muse Wonder model family
Meta released Muse Wonder and Muse Wonder Full, describing them as its first frontier models. Muse Wonder will power the Muse assistant and is available through the Meta Model API on a per-token basis; Muse Wonder Full is limited to Muse Code subscribers and is not available through the API. Capability and evaluation claims come from Meta.
**AI take:** Access differs by use case and price model, with Wonder Full limited to Muse Code.
## More News
### NVIDIA and Cisco introduce pluggable optics for AI factories
NVIDIA and Cisco announced a new pluggable-optics form factor for connecting switches and compute equipment in dense AI-factory networks. The companies say the design can help scale AI clusters; the announcement describes an architecture and product direction, not independently verified performance.
**AI take:** Optical links can affect cluster scaling, but deployment and interoperability will determine the practical gains.
### Anthropic moves some high-risk AI tests offline
Anthropic moved a high-risk model test involving cyberattack scenarios from live government websites to an offline environment. The report says models in earlier tests had accessed some government sites; the exercise was an evaluation and does not establish a real-world breach.
**AI take:** Moving tests offline reduces exposure of real systems and underscores the need for isolated evaluations.
### Gaon Cable lands a busduct contract for a US AI data center
South Korea’s Gaon Cable said its US subsidiary LSCUS signed a busduct supply contract worth KRW 200 billion (US$150 million) for an AI cloud data-center project in the United States, with deliveries scheduled throughout 2027.
**AI take:** The order shows AI data-center expansion is also creating demand for large-scale power-distribution equipment.
### KT plans a December launch for its “All of AI” service
South Korea's KT plans to launch “모두의 AI” (“All of AI”) in December, connecting domestic AI models with everyday services such as search, education and finance. KT discussed preparations at its October 12 Connect Summit; reports differ on the number of participants, and the service is not yet available.
**AI take:** The service remains a plan; its value will depend on the features and partnerships available at launch.
### Microsoft announces Maia 300 and a multi-gigawatt deployment plan
Microsoft announced Maia 300, its next-generation AI inference accelerator, and said Azure Maia deployments are planned to reach multi-gigawatt scale from 2027. The company claims Maia 300 racks can deliver up to seven times Maia 200 token throughput; this is a vendor figure, and Maia 300 deployment is still in the future.
**AI take:** The custom accelerator roadmap could shape Azure inference capacity, but the performance figure is not an independent test.
### Ambi launches a background-listening AI app after raising $26 million
Mistlabs launched its Ambi AI app and raised $26 million in a Pre-A round. TechCrunch reports that the app for Apple devices can listen in the background and use context to offer assistance; dedicated hardware remains a future plan, and persistent listening creates a privacy tradeoff.
**AI take:** Continuous context may enable more proactive help, but users must weigh that against ambient audio collection.
### Naver plans to bring Map navigation to AI glasses
Naver plans to extend Naver Map navigation to AI glasses through work with Meta, Google and Samsung. The report says Meta Ray-Ban and Oakley glasses are first, with Google and Samsung integrations to follow; not every partnership is already live.
**AI take:** Navigation could be a practical glasses use case, but rollout timing differs by device partner.
### Ondas and Palantir form the defense-focused venture Regna AI
Ondas announced Regna AI, powered by Palantir, to pursue sovereign AI for defense and named Ondas as its first customer. The companies expect initial deployments in coming months; the announcement does not show that systems are already operating.
**AI take:** Defense work puts a premium on data sovereignty, but the venture and deployments remain at an early stage.
### Ultrahuman and Abu Dhabi health body plan a biological-age model
Smart-ring maker Ultrahuman signed a two-year MOU with Abu Dhabi Public Health Center to develop and validate a biological-age model using clinical biomarkers and wearable data. The model has not yet been validated or launched; access through UAE public-health platforms is a future possibility.
**AI take:** The project brings wearable data into public-health research, with validation and data governance still ahead.
### UK AI Safety Institute gains access to test the latest models
The UK government says its AI Safety Institute can access the latest advanced models and is working with US authorities and developers including Google on evaluations. A minister said testing includes Gemini 4 Argon; this is evaluation access, not a public model release.
**AI take:** Direct testing access can move safety evaluation earlier in the deployment cycle.
### Xpeng opens invitation registration for Yoyo robotaxis in Guangzhou
Xpeng opened invitation-only registration for its Yoyo robotaxi service in parts of Guangzhou. Xpeng provides vehicles and technology while external partners operate the fleet; the limited access does not amount to a broad public launch.
**AI take:** Partner-run fleets may help operations scale, but wider access depends on local rollout.
### AMD introduces Helion, a rack-scale agentic AI platform
AMD introduced Helion, an open blueprint for an agentic-AI rack. It says customers can tailor CPU-core density, performance and power, choosing CPU-only or integrated CPU/GPU configurations; the design also includes a token router for agent calls across frontier and open models.
**AI take:** The open blueprint offers configurations for different scales of agent workloads.
### OutSystems launches Agent Experience for coding-agent governance
OutSystems formally launched Agent Experience at its World Tour event in Las Vegas, supporting coding agents including Claude Code, Cursor, Codex and Kiro within its platform. The company says security, automated testing and governance guardrails apply from the design stage, with cloud, hybrid and on-premises deployment options.
**AI take:** The product brings multiple coding agents into one enterprise governance environment; deployment will show how well it works.
### Incheon Transit wins an award for its in-house AI maintenance platform
South Korea’s Incheon Transit developed an AI mobile maintenance platform called Vehicle Doctor and won first prize in an Incheon City AI ideas contest. The operator says it can cut time spent finding fault-related materials by up to 90% and saved about KRW 287 million by developing it in-house; these results are the operator’s own account.
**AI take:** On-device processing and faster reference searches target maintenance response, while the reported gains come from the operator.
### NVIDIA contributes Vera Rubin rack designs to the Open Compute Project
At the OCP Global Summit, NVIDIA contributed Vera Rubin rack and compute-tray designs to the Open Compute Project through its open MGX architecture, giving system builders a common basis for developing next-generation AI-factory infrastructure.
**AI take:** Shared designs give partners a common rack specification to build around.
### US Air Force contracts for WindBorne balloon weather data
The US Air Force will buy data from WindBorne's autonomous weather balloons under a contract with a ceiling of $49 million; TechCrunch reported that $1.7 million had been ordered so far. WindBorne feeds the observations into its AI weather model, and the ceiling is not the amount already paid.
**AI take:** The data contract supplies observations for AI forecasting, while orders so far remain well below the ceiling.
### Shanghai model startup MOSI raises more than RMB 1 billion in Pre-A funding
Chinese technology outlet Zhidongxi reported that Shanghai-based model startup MOSI closed a Pre-A round of more than RMB 1 billion, led by Huakong. This report is limited to the financing amount and lead investor; it does not add unverified product or business metrics.
**AI take:** The funding gives the model company capital for development and commercialization, whose results remain to be seen.
### AI infrastructure company Infinigence completes a RMB 2.3 billion Series C
Phoenix Finance reported that AI infrastructure company Infinigence completed a RMB 2.3 billion Series C, bringing reported cumulative funding to RMB 4.3 billion. IPO references in the article are speculation, not an announced listing transaction.
**AI take:** The round can fund infrastructure growth, but IPO speculation should remain separate from the confirmed financing.
### HCLTech and Anthropic form a global enterprise AI partnership
HCLTech and Anthropic announced a global partnership with a dedicated business unit and center of excellence. HCLSoftware plans to integrate Claude and Claude Code into selected products; the announcement describes plans and does not establish that integrations have shipped.
**AI take:** The partnership may extend Claude into enterprise software, with usable features still to come.
### OpenAI says an Iran-linked influence campaign used AI-generated fake journalists
OpenAI said it disrupted an Iran-origin influence operation that used AI to generate social-media comments, create at least seven fictitious journalist personas and pitch long-form articles to media outlets. The attribution and scale are based on OpenAI's disclosure, not presented as an independent investigation in the report.
**AI take:** Generative tools lower the effort needed to fabricate personas and content, while platform claims remain distinct from independent verification.
## Rumors
