What's new in DevEx - December 4

re:Invent is happening this week, and AWS has made some major announcements. Biggest one is Nova 2 — a family of multimodal models optimized for reasoning and real-time workflows. Nova Forge lets enterprises train custom “Novellas,” while Bedrock AgentCore adds memory and policy controls for long-running agents.

Europe’s underdog introduced Large 3, its new flagship model built for high-end coding, analysis, and multimodal reasoning. Its main purpose is to power complex agent workflows without locking teams into a single cloud provider. Real message is: smaller, well-tuned models can outperform giant models on real business tasks.

That’s where the new Ministral 3 family comes in: a set of lightweight models designed to run locally on everyday hardware. They ship in three variants, giving developers options for raw modeling, assistant-style behavior, or logic-heavy workflows. It’s is already emerging as developers’ top choice for on-device AI.

Replit announced a multi-year partnership with Google Cloud to power its IDE with Gemini 3. This brings multi-file reasoning, code planning, and agentic edits directly into the editor with far fewer context switches. It marks one of the strongest moves yet toward a fully agent-driven coding environment for everyday developers.

Black Forest Labs released FLUX.2, a high-resolution image model with multi-reference editing and FP8 support. Its open weights mean teams can fine-tune it without waiting for API feature parity or usage caps. Shortly after, the company announced 300 million in new funding to build the future of visual intelligence.

Researchers introduced a “just-in-time” memory system that compiles relevant context at runtime instead of relying on lossy summaries. This reduces drift and boosts stability in long-horizon decision workflows. If adopted widely, it could reshape how production agents structure and store internal state.

A new approach lets models “think” using continuous visual tokens rather than converting everything into text. CoVT boosts performance on spatial and geometric reasoning tasks where VLMs traditionally break down. It’s a major signal that future reasoning models may mix visual and textual chains-of-thought by default.

New paper shows that LLMs often explain the right rule but fail to apply it — a structural gap dubbed computational split-brain syndrome. It reinforces a growing view: current LLMs are powerful pattern machines, not true reasoning systems, and will require new mechanisms for grounded computation.

Supertonic, a 66M open-source TTS model, hit breakout adoption this week. It runs 100× real-time entirely in the browser, letting teams ship speech features without latency, API keys, or cloud dependencies. Expect to see this become the default TTS choice for privacy-first or offline-focused apps.

Snapchat released Valdi, its long-standing internal UI framework that compiles TypeScript directly into native iOS and Android views. It avoids the JS bridge entirely, offering faster performance than React Native and a simpler mental model than Flutter. Teams are already experimenting with it for high-performance mobile apps.

A fast, visual Rust-based git history viewer that surged in popularity for its clean UI and speed. Terminal-heavy engineers are adopting it as a modern alternative to traditional git log workflows. Its traction shows there’s still room to reinvent core developer tools with better ergonomics.

When Andrej Karpathy releases something new, it matters (a lot). LLM Council lets developers query multiple LLMs (GPT-4o, Claude, Gemini, Mistral, etc.) at once and have a “chair” model synthesize a final, consensus answer.

The result is dramatically reduced hallucination, more stable reasoning, and a clearer view into how different models disagree on hard problems.

The tool went viral in a matter of hours, and teams are already using it for debugging, architecture reviews, and verification workflows where a single model’s answer isn’t trustworthy enough. Definitely worth a look.

Got a burning engineering question? Just hit reply and we’ll tackle it. And if you enjoyed this issue, consider sharing it with your teammates.

Want to share your dev tool, research drops, or hot takes?

Submit your story here — we review every submission and highlight the best in future issues!

Till next time,