Fish Audio raised a $52 million seed round for its AI voice platform, and the numbers suggest something’s working. The company says it’s already doing $21 million in annual recurring revenue and has 8 million people using either the open-source or hosted version of its voice models.

The Beijing-based startup launched just last year. It makes text-to-speech and voice cloning models that run both as open-source code and as a paid API. That dual approach seems to be paying off. Give developers something they can run themselves, and some of them will eventually pay for the hosted version.

The pitch is pretty straightforward: creators and enterprises need realistic AI voices, and Fish Audio is selling access to models that can clone voices or generate new ones. The company didn’t share much about pricing or which tier of customers is driving that revenue, but $21M ARR less than two years in is real traction.

This is part of a broader pattern. Voice AI went from “interesting demo” to “actual product category” faster than most people expected. There’s now a whole market of companies selling similar tools: ElevenLabs, PlayHT, Resemble AI, and others. Fish Audio’s angle is the open-source option, which matters to developers who want to self-host or modify the models.

The seed round was led by Baidu Ventures and Matrix Partners China. That’s a big seed round, even by 2026 standards, and it suggests investors think the voice AI market still has room for multiple players.

Also worth watching: Google’s agent framework gets real features

Google updated its Managed Agents API with some actually useful developer tools. The new version adds Gemini 3.6 Flash support, which is Google’s faster, cheaper model. More importantly, it adds hooks.

Hooks let you intercept and modify what an agent does before it executes. That’s critical if you’re trying to build something production-ready. You need to be able to say “wait, don’t make that API call yet” or “let me validate this action first.” Without hooks, agents are just autonomous scripts you can’t really control.

The update also improves streaming responses and adds better support for tool calling. These are the kinds of incremental features that matter more than new model releases if you’re actually trying to ship an agent-based product.

If you’re building on Google’s AI stack, this is worth checking out. The Managed Agents framework is Google’s answer to LangChain and similar orchestration tools, but with tighter integration into the Gemini API. Whether that integration is worth the lock-in depends on how much you’re already using Google’s infrastructure.

And one more thing: uv 0.12.0 changes project structure

This one’s not AI, but if you write Python, it affects you. The latest version of uv, Astral’s Rust-based Python package manager, changes the default project layout when you run uv init.

The big shift: new projects now use a src/ directory structure by default instead of putting code in the root. So instead of my_project/main.py, you get my_project/src/my_project/main.py. This matches what a lot of Python developers already do manually, but now it’s baked into the tool.

This matters because uv is becoming the de facto Python toolchain for a lot of developers. It’s faster than pip, handles virtual environments better, and keeps adding features that make Python packaging less painful. When uv changes defaults, it shifts what “normal” looks like for new Python projects.

There are other changes too: better dependency resolution, improved lock files, and some breaking changes to how the tool handles certain edge cases. If you’re using uv, read the release notes before upgrading.