This October 13 AI briefing rounds up five developments from the past 24 hours: Meta launched its first frontier model, Muse Wonder; Microsoft unveiled its third-generation in-house inference chip, Maia 300; ModelBest open-sourced a multimodal model that runs smoothly on phones; Qwen put three AI wearables on sale the same day; and Character.AI filed for bankruptcy and started an asset sale.
Meta releases Muse Wonder: its first frontier model goes closed and paid
Meta Superintelligence Labs released the Muse Wonder family on October 12, 2026, in two versions: standard Muse Wonder and the stronger Muse Wonder Full. The standard version is live on the Meta Model API and the Muse Code coding tool at $4 per million input tokens and $20 per million output tokens, with a context window of about 1 million tokens; Full is not on the API and is reserved for Muse Code paid subscribers. Unlike the open-weight Llama line, this generation is closed. Meta has formally joined the closed frontier race, and teams choosing models on that tier should benchmark it on their own real tasks before switching.
Microsoft Maia 300: 288GB of HBM4, production deployment only in 2027
Microsoft announced its third-generation in-house AI inference chip, Maia 300, on its official blog on October 12, 2026, and said it will scale the Maia family to multi-gigawatt deployments from 2027, aiming to make it the primary inference system for Microsoft products and services. Each Maia 300 carries eight HBM4 stacks for 288GB of capacity and 18TB/s of bandwidth at a 950W power budget; Microsoft claims 40% more tokens per megawatt than leading GPU systems and up to seven times the rack-scale token throughput of Maia 200 in a 96-chip rack. Both figures are Microsoft's own and not yet independently verified, and production deployment waits until 2027, so Azure users will not feel a change in the short term.
ModelBest MiniCPM-V 4.7: a 35B MoE model in about 3GB of memory
ModelBest released its on-device multimodal model MiniCPM-V 4.7 on October 13, 2026, with the official OpenBMB repository updated the same day. It ships in 35B-A3B and 1B sizes: the larger one totals 35 billion parameters while activating only about 3 billion per inference, making it the first open-source MoE model to run smoothly and losslessly on-device. Using the company's LES look-ahead expert scheduling, it reaches 13+ tokens per second on an iPhone 17 Pro Max within about 3GB of memory. It supports iOS, Android and HarmonyOS across Rockchip, NVIDIA, Intel and Apple silicon. For developers building local document understanding or offline image reading, on-device multimodal just gained another downloadable option.
Qwen AI glasses and earbuds go on sale: the N1 Pro pays by iris
Qwen officially put three wearables on sale on October 13, 2026: the N1 and N1 Pro AI glasses and an AI clip-on earbud. The N1 series carries a 50-megapixel imaging system with ambient awareness, translation and meeting notes; the N1 Pro adds eye tracking, heart-rate and blood-oxygen monitoring, and iris-verified payments completed by looking at a payment code and nodding to confirm. The earbuds are tuned with Bose and translate 89 languages. After national subsidies and launch discounts in China, the glasses start at 2,218 yuan and the earbuds at 1,099 yuan, prices that apply only to eligible buyers in the Chinese market. Qwen positions wearables as the entry point to a personal agent; how well that works awaits real feedback after launch.
Character.AI files for bankruptcy: Disney leads bidding at $20 million
Reuters reported on October 12, 2026 that Character Technologies, the operator of Character.AI, has filed for Chapter 11 bankruptcy protection in the US and begun a court-supervised asset sale. Disney is the lead bidder at $20 million for assets including a non-exclusive license to certain intellectual property, while other buyers may still bid higher; the company expects the process to conclude in late November or early December. Court filings list assets and liabilities each between $50 million and $100 million. The company says the platform will keep operating with minimal disruption during the sale, but if no buyer willing to run it emerges, the product will be wound down, so current users should watch for further announcements.
