Today brought a new model with almost every piece of news, but the one that caught my attention most was the smallest of all. A model that runs directly in your pocket without relying on the cloud. I have made it the focus of this edition, followed by the rest of what I filtered below.
Whenever I discuss with the team where to run an AI model, the question that comes up most often is: does the customer’s data need to leave their device?
This week, that question became easier to answer.
PrismML launched Bonsai 27B, a 27-billion-parameter model that runs directly on an iPhone while retaining around 90% of the original model’s quality. Nothing is sent to the cloud.
That changes the equation for anyone building credit and financial products.
In my day-to-day work with credit products, sensitive data is the rule, not the exception: receivables, revenue, and payment behavior. Every call to a cloud model is another point to consider security, cost, and compliance.
With the model running locally, that equation changes. Zero marginal cost per interaction. It works offline. The data never leaves the device.
I do not think everything will become on-device tomorrow. But for sensitive functions, such as document triage or decision support in the field, having that option on the table is already different from not having it.
Increasingly, a good product is also about choosing where intelligence runs.
For anyone curious, I have included the news link here: Read more
The rest of the radar
GPT-5.6, OpenAI’s next generation of models — Redefines the cost-benefit benchmark for coding and knowledge work, putting pressure on roadmaps that compete with Claude and Gemini. Read more
Muse Spark 1.1, Meta’s multimodal agentic model — Meta is making a strong move into agentic models with a public API, expanding vendor options for agent features and computer use. Read more
OpenAI’s new voice models for real-time conversations — Two-way voice opens the way for live translation, assistants, and more natural conversational support. Read more
Agnost AI (YC S26): extracting feedback from agent conversations — Turns real AI-agent conversations into signals for bugs, feature requests, and churn risk—almost a substitute for evals that miss real-world failures. Read more
Codex begins encrypting prompts between agents — A security change in a widely adopted coding-agent tool may affect debugging and observability for teams building on it. Read more
Cursor zero-day disclosed after the vendor failed to respond — A vulnerability in one of the leading AI IDEs is a security risk worth monitoring for anyone integrating or recommending the tool. Read more
Ask Ad Manager: Google’s AI agent inside Ad Manager — A concrete example of an agent embedded in a B2B product to accelerate insight and decision-making, a pattern that can be replicated in other verticals. Read more
MCP guide for managers — Explains the protocol and how to evaluate governance, security, and adoption when connecting AI agents to enterprise systems. Read more
Microsoft limits access to data that feeds third-party agents — Signals lock-in risk for those who depend on integrations with major platforms and reinforces the need for a contingency plan. Read more
That is all for today. Tomorrow’s radar is already being curated.