Models

Baseten Lands Four Open Models in Microsoft Foundry as First-Party Preview

Baseten's inference infrastructure enters Microsoft Foundry preview on Nov 17, 2026, serving DeepSeek V4.1 Flash, GLM-5.3 and Kimi K3 with MACC-eligible billing and GA slated for early 2027.

Baseten Brings Four Open Models to Microsoft Foundry Preview
Baseten Brings Four Open Models to Microsoft Foundry Previewblickpixel / pixabay
By James Calloway4 min read

Why it matters

  • Baseten's inference infrastructure entered Microsoft Foundry public preview on November 17, 2026, as a first-party service.
  • Launch lineup: DeepSeek V4.1 Flash, GLM-5.3, GLM-5.3 Flash and Kimi K3, available after administrator opt-in.
  • 100% of preview spending counts toward customers' Microsoft Azure Consumption Commitment (MACC).
  • Baseten targets general availability in early 2027 and plans day-one availability for new open models.
  • Infrastructure is designed for 99.9% uptime; requests are stateless and Baseten stores no prompts or responses.

Baseten's inference infrastructure entered public preview through Microsoft Foundry on November 17, 2026, powering four open models for Azure customers as a first-party Microsoft service. The launch lineup consists of DeepSeek V4.1 Flash, GLM-5.3, GLM-5.3 Flash and Kimi K3. General availability is expected in early 2027, with Baseten promising to add new open models on their release day.

The deal matters because it moves leading open-weight models into Microsoft's enterprise procurement and governance stack. Azure customers can now run these systems under their existing Microsoft agreements, with usage billed as first-party Foundry consumption rather than as a separate vendor invoice.

What does the preview actually deliver?

The four models sit in the Foundry catalog alongside other open and closed systems. All Foundry customers can use the preview, but an administrator must opt in before an organization gains access — a deliberate governance gate, not a default-on rollout. The announcement names coding agents, research workflows and customer support applications among the intended uses.

Baseten positions the integration as a way to evaluate and adopt open models without sourcing GPU capacity or managing model runtimes. The company handles the serving infrastructure. Developers choose between two deployment modes: Standard and Provisioned, both of which Baseten says are ready in seconds.

That speed claim carries the usual caveat: it refers to provisioning the deployment, not to sustained throughput under production load, which the preview period will test.

How is it billed?

Pricing splits along deployment lines. Global Standard and Data Zone Standard deployments in the US use pay-per-token billing. Global Provisioned and Data Zone Provisioned deployments in the US use provisioned throughput units (PTUs), which reserve capacity and provide predictable throughput for workloads with steady demand.

Baseten says 100% of spending through the preview counts toward a customer's Microsoft Azure Consumption Commitment (MACC). That detail ties model usage directly to an existing Azure spending arrangement and gives enterprises a financial reason to route open-model inference through Foundry rather than through a standalone API provider.

For buyers, the practical takeaway is simple: open-model inference now draws down the same commitment pool as the rest of their Azure spend.

Which APIs does it support?

Models are served through the Chat Completions API and the Responses API. Baseten says that support lets existing OpenAI-compatible code and agent workflows carry over with minimal changes, reducing the integration work needed to test the new catalog entries.

This compatibility choice lowers the switching cost that has historically kept enterprises on a single vendor's stack. A team already calling OpenAI-style endpoints can point the same code at DeepSeek V4.1 Flash or Kimi K3 and compare results before committing.

The verification burden, however, sits with the customer: "minimal changes" is Baseten's characterization, and agent frameworks with vendor-specific features may still need adjustment.

What about security and data handling?

The service runs on the same performance research and inference infrastructure behind Baseten's Model APIs. Foundry contributes the Azure security, observability and billing framework around it.

Baseten says every model is tuned across routing, batching, caching and kernels, and that the infrastructure is designed for 99.9% uptime. That is a design target, not a measured service-level figure — enterprises negotiating production agreements will want it in writing.

Data handling is comparatively restrictive:

  • Requests are stateless.
  • Baseten stores neither prompts nor responses.
  • Authentication options through Foundry include Microsoft Entra ID and key-based access.
  • Azure Policy supplies per-model access controls.

The combination of no prompt storage and Entra ID authentication addresses the two questions enterprise security teams ask first about any new AI endpoint.

What did Microsoft say?

Steve Sweetman, VP of Product at Microsoft Foundry, said: "Bringing Baseten to Microsoft Foundry gives enterprises more choice and pairs Baseten's optimized and reliable inference with the governance, controls and observability within the Foundry platform so developers can turn their most ambitious ideas into production applications."

The quote frames the partnership as a choice-and-governance play rather than a performance claim. Notably absent from the announcement: benchmark comparisons against closed models, latency figures, or pricing per million tokens — the numbers buyers would need to make an apples-to-apples decision today.

What comes next?

Public preview is the first phase. Baseten expects general availability in early 2027 and plans to keep adding open models as they are released, including making new systems available on their release day. Day-one availability would shorten the gap that has often left Azure customers waiting weeks or months for new open models to reach enterprise-approved catalogs.

For Azure teams, the immediate change is operational: open models can be evaluated through Foundry's catalog, controls and billing without building a GPU serving environment. The preview is, in effect, a live test of whether integrated infrastructure and procurement can ease the path from model evaluation to production for enterprise AI teams — and whether Baseten's day-one model pledge holds once the next major release arrives.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. Baseten opens Serverless Loops for open-model post-training
  2. OpenAI Ships o1 to Developers With 60% Cheaper Audio
  3. OpenAI frontier models and Codex land on AWS Bedrock
  4. Perplexity’s Orchestrator retries failed requests across providers
  5. OpenAI releases GPT-4.5 research preview, its largest model

« Previous articleNext article »