The AI log, frontier · robotics · image, video, world
19 Nov 2026
Microsoft's MAI lineup, Cosmos Data Factory, Comfy Builds
Microsoft ships four more MAI models, NVIDIA packages synthetic data blueprints for physical AI, and Comfy Cloud runs your own ComfyUI setup.
By Ali Asif, published
What shipped
- Microsoft AI
Seven new models at scalable price-performance
Beyond MAI Code 2, the Ignite post adds MAI Cyber 2 Flash, MAI Image 2.7, MAI Realtime 1 and the routing model Decision 1.1, each with vendor-reported price and benchmark claims.
- NVIDIA
Build Training-Ready Physical AI Datasets with Two Prompts in NVIDIA Cosmos
Cosmos Data Factory bundles five blueprints and six agent skills, and Cosmos Dataset Search is now open source under Apache 2.0.
- Comfy
Your custom ComfyUI is now in Comfy Cloud
Custom Builds package a team's own nodes, models and LoRAs and run them on Comfy Cloud on every paid plan, from $1.75 to $11.23 an hour by GPU.
- Google
The Gemini app is now available on lower-memory Android devices
The app was re-architected to run on eligible Android Go devices, with Gemini Live in more than 45 languages.
Four items today, all from pages I opened and read. I have not run any of it, so every number below is the publisher's own. I already covered MAI Code 2 earlier this week, so I am skipping it here and looking at the rest of Microsoft's list.
Microsoft's other MAI models
The Ignite post lists seven models. Four are new to this log. MAI Cyber 2 Flash is a security model for agentic defense at codebase scale. Microsoft says it beats GPT-6 Astra's quality at under 2% of the cost per task on CyberGym-E2E, and that output tokens cost 86% less than MAI Cyber 1.1 Flash.
MAI Image 2.7 is the image model. Microsoft says it ranks first on Artificial Analysis for image editing and fourth for text-to-image, and that it beats GPT Image 2.6 Sunburst at one-fifth of the price. It is in Microsoft Foundry. MAI Realtime 1 is the first live speech-to-speech model from the team, starting at $0.01 per audio minute, with a claimed edge over GPT-Realtime 2.1 on external and Dynamics 365 benchmarks.
Decision 1.1 is a model that routes work, sending a task to a larger model only when needed. Microsoft says it is 50x faster and 79x cheaper than GPT-6.1 Sol and is used by 4,000 customers. Routing is the least glamorous item on the list and probably the one that moves a bill the most.
NVIDIA Cosmos Data Factory
Cosmos Data Factory is a set of five cloud blueprints and six agent skills for curating and generating training data: driving scenarios, image attributes, video augmentation, event video and defect images. Cosmos Dataset Search is now open source under Apache 2.0, and Cosmos Curator takes multicamera video, IMU and GPS inputs.
NVIDIA reports up to 60x faster dataset assembly, with a defect dataset going from two months to one day, and about $2.66 per generated driving clip. The post is clear that these are compute costs on eight RTX PRO 6000 Blackwell GPUs, not data prep, storage or human review. Customer figures from Siemens, Encord and Voxel51 are cited, some marked as expected.
Comfy Cloud Custom Builds
A Build packages a team's supported custom nodes, models, LoRAs and dependencies, including private node packs, and runs them on Comfy Cloud. It is on every paid plan. GPUs run from RTX PRO 6000 to B200, at $1.75 to $11.23 an hour billed by the second. Concurrency is 2 jobs on Standard and Creator and up to 20 on Team. Check the supported-nodes list before you package anything.
Gemini on lower-memory Android
Google re-architected the Gemini app so it runs on eligible Android Go devices. It brings Gemini Live in more than 45 languages, camera translation of notes and forms, and image creation. The post names no memory threshold or device list.
What I'd watch next
Independent runs of MAI Cyber 2 Flash against the cost claim. Whether Cosmos defect datasets hold up once human review is counted. And which Go devices ship Gemini preloaded.
Sources
Topics: NVIDIA, Gemini, Google, Microsoft, ComfyUI, Cosmos, Microsoft AI