DeepSeek open-sources engine behind its cheap AI

Welcome back. For two years, U.S. labs have raced to match DeepSeek on price and performance. Now DeepSeek has open-sourced its inference framework for anyone to run. Meanwhile, two fresh voice agents just rolled out, both eager to power your workflows. 

Also: A reminder of the real reason LLMs code so well, a viral post highlights which category of engineers AI could replace first, and how a founder puts his agent to work on online haters.

DeepSeek open-sources the engine behind its affordable AI: The Chinese AI lab just dropped DeepInfer, the software it uses to run V4.1-Flash. If you host the model on supported Nvidia or Huawei chips, you can now use DeepSeek’s production stack to run it faster and cheaper. Even if you use another model, you can still borrow ideas from the code for your setup. It’s not the full system, though, so you’ll need to build a few pieces yourself. Grab the repo.

Cua lets your agents run jobs across every machine: The computer-use startup just shipped its Spaces SDK, giving each of your AI agents a desktop with its own screen and apps. Unlike Grok Bot and Muse, you get to choose where they work, and moving a job to another machine takes just a one-word code change. Claude Code, Codex, and OpenCode can all use the same space and tools, too. Try it here.

ElevenLabs and Mistral drop new models for voice agents: Your voice stack just got two new contenders. ElevenLabs’ new speech-to-speech models, Eleven Realtime, let you switch from OpenAI’s Realtime API by changing only the endpoint and model name. The cheaper Mini model handles natural chats, while Max reasons through complex tasks. Mistral’s Voxtral 3 gives you more room to tinker. Its LiveKit-based SDK lets you swap the LLM, voice, or language without rewriting your agent. With Mistral’s help, you can also run the whole stack on your own servers.

Most voice AI still works best in English.

Soniox brings speech-to-text, text-to-speech, and translation together in one real-time platform across 60+ languages, built for accurate input, natural output, and seamless multilingual conversation.

Powering real voice products at Perplexity, Samsung, Wispr Flow, Vapi, Retell, and more.

Build with Soniox

Ten years overnight. Uber's AI coding tools might look like an overnight win, but building them took a decade. On a recent podcast with veteran engineer Gergely Orosz, Distinguished Engineer Uday Medisetty explained how all that preparation finally paid off. Today, Uber's background coding agent, Minion, ships roughly 15% of the company's production PRs. Keep in mind, engineers built much of that essential foundation long before LLMs even existed. 

Old stack, new tricks. Minion runs on DevPods, the cloud dev boxes Uber started building in 2017. A DevPod spins up in seconds, with repos cloned, build caches warm, and AI tools installed. The MCP Gateway came together quickly for a similar reason. Uber already kept service API contracts in one registry, so it could automatically turn those contracts into agent tools.

Check yourself. An agent that writes code also needs a way to test it. At Uber, agents deploy their branches to Slate, a private staging setup that routes a test account through the changed code. Even that relies on older infrastructure, Jaeger tracing. Without Slate, Medisetty says, agents would have had a much harder time validating their work.

Boring pays off. You probably don’t need Uber’s entire platform, but you can borrow the groundwork:

Fix the first layer that fails. Orosz's key takeaway is that the companies benefiting the most from AI already had dedicated developer platform teams in place, which was certainly true for Uber. In addition, the episode post maps out each system alongside detailed timestamps. 

How to make your agent improve its own harness: This tutorial shows you how to get an agent learning from real conversations, spotting where its rules, skills, or memory fell short, and proposing fixes to its own setup. You stay in control of what actually changes, while the agent handles the tedious part of figuring out what needs improving.

Firecrawl (193.6k ⭐): Give your agents a much better way to use the web. Firecrawl can search, scrape, crawl, click around pages, parse PDFs and DOCX files, and return clean Markdown or structured JSON instead of messy raw HTML. You can plug it into Claude Code, OpenCode, Antigravity, or any MCP client in a few commands.

How top Silicon Valley engineers build software factories: Our in-house cookbook pulls together lessons from engineers at OpenAI, Anthropic, Stripe, and Warp to help you build one from scratch, then make it faster, cheaper, safer, and easier to scale across a team. You’ll cover setup, cost controls, evaluation, and governance.

A developer shared a project rule that fixes a common vibe-coding annoyance. AI-built dashboards render stats as dead numbers with no trend, no context. One paste fixes every future screen.

Vercel’s CEO says the 9-to-5 is gone for AI founders. What they replaced it with sounds even more brutal.

If you enjoyed this newsletter, you should definitely check out our IT and cybersecurity newsletters. Subscribe using the links below:

We put together 20+ hands-on guides covering coding agents, agentic engineering, and leadership frameworks to help you level up your career. You can get them here.

Advertise with us: Join companies like Google, IBM, and Datadog. Showcase your product to our 350K+ engineers and 150K+ followers on socials. Get in touch.

You can also reply directly to this email if you have suggestions, feedback, or questions.

Until next time — The Code team