Baseten

The voice pod at Baseten has been shipping behind the scenes but today we're launching them all: the best whisper + speaker diarization stack. How is it the best? https://lnkd.in/eU_mDGZN, but in short: • 2400× real-time factor with Whisper Large V3 Turbo • Streaming transcription with consistent low latency under load • The most accurate real-time diarization (see screenshot) • ~90% lower cost than other providers due to infra optimizations The key wasn’t a single model tweak; it was decoupling the system using Baseten Chains, but more on that in the blog post. This stack now powers Notion's AI Meeting Notes with live, speaker-aware transcription at scale. And give AI Meeting Notes a shot; it's replaced others for most folks at Baseten.

"Engineers who are product-minded and know what they want to ship, because they understand their users, now they have infinite power." Madison from Baseten sits down with the Sourcegraph team to discuss all things code gen and coding agents. Full vid ⤵️

Inference performance comes from much more than the model. It requires runtimes and infrastructure optimized from the ground up—an entire inference stack.. In our Inference Stack white paper, we break down how Baseten leverages tools like NVIDIA TensorRT LLM and NVIDIA Dynamo as part of our ecosystem to: > Optimize custom kernels > Maximize throughput for LLM, embedding, voice, and vision workloads > Minimize latency through intelligent routing and KV cache optimizations If you care about speed, this is worth reading: https://lnkd.in/ev49FtRK

"It's going to let you take the world as you see it, or let the code that you have that works today [...] and just go." Check out the inside story on Baseten's newest product ⤵️

Here’s the part people don’t like to say out loud: in most practical settings, RL is the wrong starting point. The criteria to choose between RL vs. SFT for training your own models is not what you think. Check out pt 2 of our series ⤵️ https://lnkd.in/eZy4H4Wh

2025 redefined what it takes to get an AI application production-ready. Here are 3 infra learnings that will help you build better in 2026: 1. Reliability is the new moat 2. Speed is the product 3. Developers want the wheel Read the full Baseten AI Wrapped in the comments ⤵️

Alex Ker from Baseten recently spoke at AI Dev x NYC by DeepLearning.AI. In this talk, he covers the rise of open-source coding models and how you can take advantage of these OSS models in your current dev workflows in less than 10 minutes Watch the full video ⤵️

Turns out we know a lot of people that like to go fast. Congrats to Stewart Noll and Alex Choy for taking the top two spots at our need for speed event at Re:Invent! A big thank you to our friends at Graphite for co-hosting with us, and to all our friends from Notion, Gong, Bloomberg, Air, and more for showing up ready to race. If you think you have what it takes to make it on the leaderboard, shoot us a note!

Baseten launches day-zero support for NVIDIA Nemotron 3 Nano Nemotron 3 Nano delivers: up to 4× faster token generation, high accuracy, and predictable inference costs, making it ideal for agentic AI in financial services, retail, and beyond. Deploy today on Baseten for low-latency, high-throughput inference powered by the Baseten Inference Stack and NVIDIA. Learn more here: https://lnkd.in/eTVdV7aJ

RL is a critical step towards serving production-ready AI applications with very specific needs requirements. In this guide, we discuss why reinforcement learning matters and break down RL vs SFT in practice. Check out the full guide in the comments ⤵️