Open enterprise models, world models, and concrete agent tools for your next sprint

Important AI news

NVIDIA Kumo: Open foundation models for structured data

NVIDIA introduces Kumo with open foundation models, libraries, and tools for predictions from structured enterprise data. That covers tasks like forecasting, anomalies, and risk scoring. These often need their own ML pipelines. Early users include Cognite, Coinbase, Databricks, DoorDash, Revolut, and Tubi. The announcement came last week in Berlin. The listed performance numbers come from the vendor.¹

Cosmos Data Factory and framework: More speed for world models

NVIDIA expands Cosmos with data factory blueprints and training tools. A world model is a model that understands sensor and video data so it can simulate the physical world and test actions. Five cloud blueprints cover driving scenes, image attributes, event videos, defects, and video augmentation. They are available on GitHub, Azure, Nebius, and CoreWeave. Porsche expects fewer real test drives and shorter cycles. These are vendor numbers and four days old.²

Financial services deploy open models in production

Revolut is rebuilding its stack with the foundation model PRAGMA. It will first serve 80 million customers and is sized for a 3x expansion. Neobank bunq reports 60 percent lower answer costs at comparable quality for assistant Finn with open weights. It runs on Nebius. These numbers come from the vendor. More examples range from ATUM at Lloyds to faster data pipelines for fraud detection.³

Highlights for your workday

Uber rolls out ChatGPT Work to more than 30,000 employees. You can infer an internal pilot size and typical tasks from that. Research, analysis, slides, complex workflows.⁴

Amazon Quick explains the path to a HIPAA-capable agent workspace. The post cleanly separates HIPAA-eligible and compliant. It shows which protocols, logs, and rights you must configure. If you work with health data, you can plan chat agents with guardrails this way. Assumes a BAA and correct settings.⁵

Audible shows how controllable TTS makes audiobooks sound better. An LLM describes speaking styles in natural language and controls voices consistently over many hours. The tests show 3.7 points better dialog separation in the MUSHRA method and 97 percent correct speaker assignment in English. Both are vendor figures.⁶

Tools and updates

• Omnara: Open-source API for persistent agents. These are agents that keep state and survive restarts. Agents run as robust loops. State lives in Postgres. Commands run in attached sandboxes or your laptop. Self-host or use as a service.⁷
• Open WebUI 0.12.4: Lazy loading for skills, visible marker on context compression, and finer visibilities for controls. Useful if you manage many users and want to cut token load. Update to 0.12.4 required.⁸
• Amazon Trainium now closer to native PyTorch training. Eager mode, torch.compile, and .to("neuron") instead of XLA special paths. SemiAnalysis reports good direction but not CUDA maturity yet and some bugs. Suitable if you train with BF16 and want to test NVIDIA alternatives.⁹
• Replit Monitoring Agent: An on-call agent that detects, fixes, and rolls out changes for AI app teams. It builds on the Replit ecosystem and automates ops for small apps.¹⁰

Try it in five minutes

Compare reasoning settings for code and decide your team default. Own exercise idea, prompted by measurements on Claude Sonnet 5.5.¹¹

1) Set the task. A small function with clear acceptance criteria. For example: validate email and extract domain. 2) Run two passes. One with high reasoning level, one with maximum level in your chat tool. Note displayed token costs if available.¹¹ 3) Check the result. Copy both solutions into your repo, view the diff, run the linter. Watch for diff size, nested conditions, and warnings. Expectation: maximum reasoning settings do not always deliver the most mergeable code and often burn more tokens.¹¹

Result: you have a documented choice for the default setting. Check criterion: smaller diff and fewer linter warnings with accepted functionality.

A good find

Why compute is not rectangular: Modal CEO Erik Bernhardsson on elastic GPUs. Workloads fluctuate. Fixed rectangles waste budget. Pooled capacity and fast start count as real product features. Not just infrastructure details. Structure contracts so you catch peaks without idle. Measure cold start times of your inference.¹²

Sources

1. NVIDIA Open Models Advance Europe’s Sovereign Intelligence Across Industries (blogs.nvidia.com)
2. NVIDIA Cosmos Opens New Tools to Train, Evaluate and Deploy World Models (blogs.nvidia.com)
3. Leading European Financial Services Organizations Embrace Open Models to Build Proprietary Intelligence (blogs.nvidia.com)
4. Uber brings ChatGPT Work to all employees with OpenAI (openai.com)
5. Deploy AI in a HIPAA-eligible environment with Amazon Quick (aws.amazon.com)
6. Better speech synthesis for audiobooks (amazon.science)
7. Show HN: Omnara – Open-source alternative to Claude Managed Agents (news.ycombinator.com)
8. Open WebUI v0.12.4 (github.com)
9. Amazon’s Trainium Takes a Strike at the Training King (newsletter.semianalysis.com)
10. @replit: You now have an on-call engineer right in your pocket.

Meet Monitoring Agent.

Replit Monitoring Agent can see, fix, and deploy changes autonomously for AI-app builders.

Create y (x.com) 11. @deeplearningai: Max reasoning should mean best results, right? 🤔

Not always. Claude Sonnet 5.5 wrote less mergeable code at max than at xhigh, one level below, while using more tokens than any mo (x.com) 12. Why AI Compute Comes in Rectangles — Erik Bernhardsson, Modal CEO (latent.space)