AI Roundup: Cursor's Kimi-Powered Composer, OpenAI's Astral Acquisition, and Running 397B on a Laptop

The AI space keeps moving fast. Three stories dropped this week that deserve your attention if you care about coding tools, the Python ecosystem, or just how far we can push local AI.

TL;DR: Cursor Composer 2 runs on Kimi K2.5, OpenAI acquired Astral (the company behind uv, ruff, and ty), and a new project got a 397B parameter model running on a laptop.

Cursor Composer 2 Now Uses Kimi K2.5

Cursor dropped Composer 2 this week, and it's a beast. The new model shows "frontier-level coding performance" according to their benchmarks. But here's what caught everyone's eye: it's built on Kimi K2.5 from Moonshot AI.

The Kimi team confirmed the partnership, noting that Cursor accesses their model via Fireworks AI's hosted RL and inference platform as part of an authorized commercial deal. This is a big win for open model ecosystem players, seeing their models power a major commercial coding product.

Composer 2 is available now at $0.50/M input tokens and $2.50/M output tokens in standard mode, or $1.50/$7.50 for the faster option.

OpenAI Acquires Astral

In a move that sent shockwaves through the Python community, OpenAI announced it's acquiring Astral. That's the company behind uv, ruff, and ty three increasingly essential tools in the Python ecosystem.

uv is a blazing-fast Python package installer and resolver. Ruff is the popular linter written in Rust. Ty is their newer type checker. These tools have become load-bearing infrastructure for thousands of Python projects.

What this means for the community remains to be seen. Astral's team will join OpenAI, and the tools will presumably continue as open source projects. But it's a reminder that the Python ecosystem is now deeply entangled with AI companies.

Running a 397B Model on Your Laptop

This one is wild. A new project called Flash-MoE figured out how to run a 397 billion parameter model on consumer hardware. The trick is Mixture of Experts (MoE) optimization, which only activates the parts of the model needed for each input.

The project is on GitHub and shows that local AI is getting more accessible. We've seen quantized versions of large models run on desktops, but a full 397B model is something else entirely.

This points to where things are heading: AI models that can run locally without sacrificing too much capability. For developers building AI-powered apps, that's a huge deal.

Sources: Cursor Blog, Kimi Twitter, Astral Blog, OpenAI Announcement, GitHub Flash-MoE