DigitalOcean

Now available: NVIDIA Nemotron™ 3 Diarization on DigitalOcean Serverless Inference. An open-weight, streaming model that tracks who spoke when, up to 8 speakers in live and recorded audio. The previous Streaming Sortformer generation handled 4. Keep the ASR you already run. Nemotron 3 pairs with Parakeet, Canary, Whisper, or whatever's doing your transcription, so you add speaker attribution without swapping models. One 100M-parameter model replaces the usual four-stage pipeline, and your audio stays on DigitalOcean instead of routing out to a third-party speech API. Start building: https://do.co/4rtI8it

Your workflow probably runs a classification, routing rule, or moderation check through a frontier LLM at frontier prices. Jev is built only for that job. TypeSafe AI's new model returns a typed decision instead of text: which queue, what score, allow or block. It answers in 70–500ms, where a frontier model takes 3 to 329 seconds. DataCamp, reading TypeSafe's published numbers, put it within a point of GPT-5.6 Terra on accuracy at about 1/76th the cost per decision. Output is schema-constrained, so it can't hand your code a malformed answer. It can still be confidently wrong, which is why every call returns a confidence score and your code sets the bar. It won't write your copy or debug your code. Put it in front of a frontier model and stop paying reasoning prices for yes/no work. Now available through DigitalOcean Serverless Inference. https://do.co/4xJjqMw

Most agent problems in production aren't model problems. They're runtime, auth, and isolation problems. Join us for the night shift after The AI Conference with Box and Bright Data. We'll dig into the layer underneath the agent: sandboxes that boot in seconds, sessions you can pause, fork, and hand to a teammate, and tool access that doesn't mean wiring OAuth flows for the eighty-seventh time. Live demos, cold drinks, and a room full of people actually shipping agents. RSVP today. https://do.co/4AqzvcA

It's time to move your AI agents to the cloud. Brittle scripts and rigid automation pipelines are on their way out. Agents with tools are now smart enough to replace all of that heavy lifting, and DigitalOcean Managed Agents helps you deploy them in minutes. Bring the harness, model, and tools you already use, and put them in the cloud to do real work under careful access controls. You own the agent, we handle running it. Look at Qencode: they built an agent to handle their customer support triage. It spins up the moment a request comes in, assesses the issue, and files the ticket perfectly. In this demo, Ryan O'Connor shows you how to build one just like it. DigitalOcean Managed Agents is now in public preview. https://do.co/4xN8zRM

Our Slack bot used to push releases. This week it wrote the code going into those releases, picked up a bug that a colleague flagged halfway through the thread, and had a tested PR open 20 minutes later. If you're building agents that run alongside internal workflows, keep an eye on what Musa and team are doing 👀

More speakers just joined Open Intelligence Summit. Coding agents, inference economics, and the products people actually open every day. This batch covers all three. Tommy Eastman, Nous Research Yifan Qiao, Inferact Peter Walker, OpenRouter Spenser Skates, Amplitude Eric Simons, bolt.new Oct 13 at The Midway in SF. Request a seat: https://do.co/4ApUvzV

We're so excited to have Nous Research, OpenHands, Inferact, and LanceDB, as well as long-time DigitalOcean builders like Yujian Tang, join us in October!

Do code sandboxes really save tokens? Staff Developer Advocate Narasimha Badrinath wanted to see the difference between running tool calls individually vs. running them in a sandbox. The task: find the 20 most recently closed issues in vuejs/core and compute the average time to close. Real repo, live data, same underlying GitHub tool in both runs. One of these approaches cost 52% less. The tool is identical in both runs. A single lookup has nothing to strip out, so call it directly. But the moment a task chains calls, works through a large result set, or has to process something before it can answer, the sandbox pays for itself. Code sandboxes are available now as part of DigitalOcean’s Action Gateway. https://do.co/4xN8zRM

An agent spends most of its time turning messy input into structured facts and making small judgements, with a few hard problems in between. Two new OpenAI models landed on DigitalOcean Serverless Inference this week, one for each kind of work. GPT-6 Luna: classification, extraction and routing at volume, or as a cheap sub-agent tier under a bigger model. GPT-6 Sol: the multi-step coding and long-running agentic work, with a 1.05M token context window. Both run on the same endpoint and bill as the agents that call them. Try them now in Inference Engine. https://do.co/4rtI8it

TypeSafe AI founder Diogo Almeida spent four years on one question: models have excelled at chat for years, but where is all the automation? The answer is Jev, now on DigitalOcean Serverless inference. It takes a piece of state and a set of questions with answers you define in advance, and returns typed values with calibrated confidence scores. Nothing to parse, no answer outside the set you allowed. The rest of the frontier spent the month shipping coordinator agents in Cursor, 800,000 lines of Rust ported by agents at GitHub, AI receptionists booking jobs, agents reconciling insurance claims. All of it is software taking action on models trained to produce responses people prefer reading. Jev skips prose entirely for decisions software can consume directly. TypeSafe calls this a "System One" model, after the fast, intuitive half of Kahneman's "Thinking, Fast and Slow." They report it runs two orders of magnitude faster and more efficiently than an LLM on the same tasks. https://do.co/4rtI8it