Baseten 推出 Grid、Sandboxes 并正式开放 Loops 训练基础设施

Please rotate your device

To continue exploring this website, please rotate your device to a vertical orientation

Resources

*   [Blog](https://www.baseten.co/blog)
*   [Guides](https://www.baseten.co/resources/type/guide)
*   [Webinars](https://www.baseten.co/resources/type/webinar)
*   [Security](https://www.baseten.co/security)
*   [Events](https://www.baseten.co/resources/type/event)
*   [Changelog](https://www.baseten.co/resources/type/changelog)
*   [Contact Us](https://www.baseten.co/talk-to-us)
*   [Careers](https://www.baseten.co/resources/careers)

Featured

Image 13: The building blocks of the open frontier: introducing Grid, Sandboxes, and Loops The building blocks of the open frontier Blog / News

Resources

*   [Blog](https://www.baseten.co/blog)
*   [Guides](https://www.baseten.co/resources/type/guide)
*   [Webinars](https://www.baseten.co/resources/type/webinar)
*   [Security](https://www.baseten.co/security)
*   [Events](https://www.baseten.co/resources/type/event)
*   [Changelog](https://www.baseten.co/resources/type/changelog)
*   [Contact Us](https://www.baseten.co/talk-to-us)
*   [Careers](https://www.baseten.co/resources/careers)

Featured

Image 15: The building blocks of the open frontier: introducing Grid, Sandboxes, and Loops The building blocks of the open frontier Blog / News

News

The building blocks of the open frontier

This week we’re expanding our inference platform with three launches alongside our fresh new website.

Authors

Image 17: Amir Haghighat Amir Haghighat

Date

November 3, 2026

Image 18: The building blocks of the open frontier: introducing Grid, Sandboxes, and Loops

Table of contents

Grid: route to the right model for every coding task

Newsletter

Subscribe to our newsletter

Submit

Video 1

Every AI product runs on a system: models, the infrastructure that serves them, the training that improves them, and the agents that orchestrate them all. Many teams stitch that system together from separate vendors, and we’ve seen the costs firsthand. They lose time connecting the pieces, add latency moving data between them, and face downtime when one piece fails.

We’ve learned a lot from our customers about what it takes to build that system well, and many of them helped shape what we’re launching this week. OpenEvidence trains models that outperform closed models on Baseten Training, then serves them directly via Dedicated Inference. Datadog trains models on Loops, and serves models via Dedicated and Model APIs. Opendoor routes every AI request across the company through Grid (launching tomorrow), sending each one to the right frontier or open model without changing the harnesses their developers already use.

Each of these use cases needs a tight loop between inference, training, and execution. We built Baseten to bring all of those pieces into one unified platform. Every piece shares the same optimized infrastructure and runs on Multi-cloud Capacity Management across dozens of clouds, delivering 99.99% uptime while making intelligence many times faster and cheaper to run at scale.

This week, we’re launching three product milestones to extend our inference platform:

Grid: our new gateway, router, and guardrail platform

Sandboxes: launching Thursday

Loops: our serverless training infrastructure, now generally available

You might notice some new visuals on our site as well; they lean more heavily into the base ten building blocks that inspired our name.

Grid: route to the right model for every coding task

Image 19: Grid: route to the right model for every coding task

We’ve seen an enormous uptick in coding agent usage, and an explosion in spend has followed. Organizational usage is sprawled across models, providers, and harnesses. Developers default to the most intelligent model available, not the one that’s best-suited to a specific task.

At the same time, open-weight models have caught up to frontier performance on coding, while costing much less to run. Companies need a centralized platform to govern model usage and route tasks to the best model for the job, not just in terms of token cost but quality and overall price. That routing also needs to maximize KV cache hit rate, so multi-turn agentic sessions keep landing on a warm cache and cost per task keeps dropping.

Baseten Grid is our single control layer that gives enterprises access, visibility, and control over their coding spend. It has three parts:

Gateway: puts every model and dollar of spend in one place

Intelligent Router: matches each task to the right model

Guardrails: set company-wide policies on who can use what, and what models are allowed to do

Grid is harness- and model-agnostic, compatible with whichever tools your developers already use (including Claude Code, Codex, OpenCode, and Pi). Each task gets matched to the right model, with the performance, reliability, and security you know from Baseten.

Grid launches tomorrow as a public preview, and we’ll release more details then.

Sandboxes: toward the cloud for agents

Image 20: Sandboxes: toward the cloud for agents

Agent performance is a systems problem. Running code in a container isn’t hard; the hard problems are storage (state), networking (millisecond-long cold starts), and security (VMs). That's what Blaxel has built for sandboxes, and why we recently acquired Blaxel to bring sandboxes to Baseten.

Now, we’re bringing together specialized compute, storage, and networking to build the cloud for agents. Sandboxes are one piece in that puzzle. Baseten Sandboxes are stateful, pause/resumable, VM-isolated, and scale to thousands of instances in seconds.

Because Sandboxes will be co-located with inference, you also get better performance, visibility, and reliability across the full loop between inference and execution.

Sandboxes are moving from private to public preview, and Paul, our Head of Agent Infrastructure, will go into more depth later this week.

Serverless Loops: train without the infrastructure management

Image 21: Serverless Loops: train without the infrastructure management

It’s become clear that companies want to own their intelligence layer. We help teams like Datadog, Harvey, Lovable, and Parallel post-train open models for their specific use cases, continuously improving on their own data to build durable intelligence.

But not everyone wants to deal with the underlying infrastructure. Top researchers working on training should spend their time on their data and reward functions instead. Loops is our answer to that: serverless SFT and RL on frontier open-weight models, so you can focus on shaping the model you need.

The Loops SDK is also Tinker-compatible, and we let customers run on single-tenant capacity. Loops is fully integrated with our Dedicated Inference, so after training on your incoming data, you can serve your model directly as a production inference endpoint. You always retain ownership of your trained weights.

Loops has been available as a public preview for some time. It will be generally available starting Friday.

Inference, training, routing, governance, and sandboxes are the building blocks of every AI product today. When they come together on a unified platform, the whole is greater than the sum of its parts. That’s the future we’re building for.

Stay tuned this week for more technical content coming about each of these products. You can join the waitlist for the products we’re launching this week here.

Products

Resources

Legal

Developer

All systems normal

Image 24: Baseten

The open frontier is built on Baseten