The week in startup news: Oct 19-25

1. AI tool catches silent agent errors

Laminar, a startup from Y Combinator's summer 2024 batch, has been building open-source (free for anyone to use and change) tools to watch over AI agents since 2024. They just released flow-1, a model trained to spot errors agents make quietly, without crashing or throwing an obvious error. It matches the detection quality of a much bigger model while costing 23 times less, which means a small team can afford to check every single agent run instead of sampling a few.

2. India health startup raises $50 million

Even Healthcare, based in India, raised $50 million in a Series B round (a later-stage funding round once a company has some proof it works) led by Khosla Ventures. The company wants to flip how Indian hospitals make money: today, hospitals often earn more the more tests and procedures they run, even unneeded ones. Even pays and rewards doctors for good patient care instead, and is now worth $300 million, three times what it was worth 18 months ago.

3. AI tests robots before they go live

Laplace Intelligence, based in Ann Arbor, builds automated red-teaming (deliberately trying to break a system to find its weak points) for robots and self-driving cars. Their bet is that the only reliable way to catch dangerous failures in physical AI is to have another AI try to break it first, in a lab, before real people are anywhere near it.

4. CERN and Mistral team up on AI

CERN, the European particle physics lab, and Mistral, a European AI company, signed a statement of intent to build foundation models (large, general-purpose AI models that can be adapted for specific tasks) for scientific research. It pairs CERN's huge physics datasets and computing power with Mistral's experience training large models.

5. Latin America startup funding up 26%

Startups in Latin America raised $1.3 billion in the third quarter of 2026, up 26% from a year earlier, according to Crunchbase. Late-stage and growth deals grew 74% year over year to $757 million, and seed-stage funding (the very first money a startup raises) also picked up, suggesting investors are starting to look at the region again.

6. Meta's new free AI model runs on Macs

Meta released Muse Spark Open, an open-weight model (one where the trained settings are free to download and run yourself) built specifically to work on Apple computers instead of needing cloud servers. It has 816 billion parameters (the internal settings that shape what the model knows) and can handle very long inputs, but you'd need 1 terabyte of memory to actually run it, so it's not something most people can use at home yet.

7. New AI model spreads fast on Vercel

Jev, made by TypeSafe AI, is a lightweight model built just to pick between a few preset options rather than generate open-ended answers. That narrow focus lets it run cheaper and faster than bigger models, and within two weeks of launch it was already used by 35% of paid teams on Vercel, the fastest adoption Vercel has ever recorded.

8. India opens defence deals to startups

Following Operation Sindoor, India's military is working more directly with private technology companies instead of relying only on state-run defence manufacturers, according to industry leaders speaking at TechSparks 2026. Founders described faster procurement and more chances for private companies to contribute critical military technology.

9. Evidence (YC S21) is hiring engineers

Evidence builds business intelligence software (tools that help companies make sense of their own data), used by thousands of teams from startups to Fortune 100 companies, often replacing tools like Tableau and Power BI. Backed by Y Combinator and with 7,000 GitHub stars on their open-source project, the small team is now hiring software engineers.

Why this matters if you're building

This week shows capital and tools moving toward smaller, more open setups. Money is flowing again into regions like Latin America and India, big companies are giving away capable AI models for free, and small, focused tools are beating larger general ones on speed and cost. If you're building with a small team, there are more open doors right now than there were a year ago, whether that's funding, free AI models, or jobs at proven early-stage companies.