What was announced
Baseten, a private AI inference company, has partnered with Microsoft to offer its open-model inference inside Microsoft Foundry, now in public preview for Azure customers. It is available today in public preview, with general availability expected in early 2027. Treat that GA date as a forecast, not a promise.
Microsoft's own Foundry blog, published during Ignite, confirms the preview. It describes the offering as bringing Baseten-optimized open models such as Kimi-K3, GLM 5.3, GLM-5.3-Flash and DeepSeek V4.1-Flash-1017 into Foundry as a Microsoft first-party service, tuned for speed, throughput and cost. The "first-party" label matters. It means Baseten's service is sold and supported through Microsoft's Foundry commercial relationship. It isn't just a link to a separate vendor's endpoint.
The model lineup
Four models are named:
- DeepSeek V4.1 Flash 1017
- GLM 5.3
- GLM 5.3 Flash
- Kimi K3
Baseten says each one runs on its full inference stack. No independent benchmarks of that stack's speed, throughput or latency accompany the announcement, and no pricing figures were published. Treat "fast" and "reliable" as vendor claims until someone measures them.
One caution on the catalog. Foundry's catalog already has a DeepSeek-V4.1-Flash-1017 entry, listed as hosted on Azure. It shows text and image input and a context window of roughly one million tokens. That entry confirms the model exists in Foundry. It doesn't prove that entry is served by Baseten, so don't assume the two are the same.
Billing and capacity: what's live and what's planned
Baseten says customers are billed through Azure as first-party Foundry consumption and can apply their existing Microsoft Azure Consumption Commitment. For finance teams, that means no new vendor onboarding and spend that counts against an existing commitment. Nobody has said it will cost less.
The press release says teams can choose pay-as-you-go or Provisioned Throughput Units (PTUs) for SLA-backed performance. Microsoft's blog is more careful about timing. It says to start with serverless pay-as-you-go, globally or in the US Data Zone, and scale with provisioned throughput starting in December.
Put plainly:
- Serverless pay-as-you-go is the preview-day option.
- Provisioned throughput is a December rollout, according to Microsoft.
Microsoft Learn describes PTUs generally as reserved processing capacity. It says minimum PTU requirements vary by model and that availability depends on deployment type and region. That is general Foundry guidance, not Baseten-specific detail. Plan capacity and check regional support before you commit.
Governance and data handling
Baseten's release lists four enterprise controls:
- Global and U.S. data zones for residency needs.
- Zero data retention, with no prompts or responses stored.
- Azure Policy controls that let administrators restrict specific models.
- Billing through the existing Microsoft agreement.
Microsoft's blog corroborates global and US Data Zone access. The sources I reviewed don't independently document the zero-retention implementation or the exact policy controls for this service. Those are announcement claims.
Microsoft Learn explains how Data Zone deployments work in general. Data Zone types process data only within the Microsoft-defined zone, while Global types may process data in any Azure region. Learn also shows an Azure Policy pattern that blocks a specific Foundry deployment type by SKU name. That is useful background for administrators. It isn't proof that the Baseten offering supports a built-in per-model block.
Before approving production use, administrators should ask these questions:
- Does the selected zone cover all inference processing, or only some of it?
- What exactly does "zero data retention" cover, including logs and telemetry?
- Can the Azure Policy you need actually be applied to this service's deployments?
- Does the preview's terms of service allow your workload?
How developers get started
According to Baseten, an enterprise first opts in. Developers can then find Baseten models in the Foundry model catalog, choose a model and pick a deployment type. The company says deployments are ready in seconds. I found no documentation of the opt-in path, role requirements or supported API surface, so I won't give a step-by-step procedure.
Why it matters
Microsoft says Baseten joins Fireworks, already generally available on Foundry. Foundry now carries multiple partner-run inference providers for open models. For Azure shops, the practical upside is straightforward:
- Open models can be evaluated in the same workflow as other Foundry models, with existing identity and governance controls.
- Spend can stay on one Azure agreement.
- Microsoft support covers it, per Microsoft's blog.
The caveats are equally clear. This is a public preview. Performance claims are unverified. Provisioned capacity isn't available yet. "Open model" here means open-weight models run by Baseten as a service. It doesn't mean you host them or own the infrastructure.
For now, it's worth a pilot on non-critical workloads, with your own evaluation data.
References
- Baseten Brings High-Performance Inference for Open Models to Microsoft Foundry - The Manila Times The Manila Times · 2026-11-17T23:22:00+00:00
- deepseek-ai--deepseek-v4.1... | Model Catalog | Microsoft Foundry ai.azure.com
- Baseten Brings High-Performance Inference for Open Models to Microsoft Foundry quantisnow.com