Subscribe
News
4 min read

Microsoft Brings 26 Fireworks-Served Open Models to Foundry

Microsoft Foundry listed 26 open models served by Fireworks AI. Startup credits cover only pay-per-token use, and EU, FedRAMP and card-data limits apply.

Olivetti-style cover reading 26 Open Models on Azure via Fireworks, with gear, code bracket and network icons in colored tiles
Olivetti-style cover reading 26 Open Models on Azure via Fireworks, with gear, code bracket and network icons in colored tiles

Microsoft is courting startups with open AI models on Azure. On August 4, 2026, the Microsoft for Startups team published a deployment blueprint for running models served by Fireworks AI inside Microsoft Foundry, Azure's platform for building AI apps. Members of the Microsoft for Startups program can apply their Startup credits to those deployments, with some limits. Unite.AI reported that the Foundry catalog then listed 26 Fireworks-served models and that the program advertises up to $150,000 in Azure credits.

For a non-developer, the short version: a startup can rent well-known open-weight models, such as DeepSeek, Kimi, Qwen and OpenAI's gpt-oss, through its normal Azure account and bill, instead of buying GPUs or signing a separate contract with an inference company.

What Fireworks on Foundry is

Fireworks AI is a company that runs open models on its own GPUs and sells access to them. Microsoft says Fireworks on Foundry became generally available on June 2, 2026, as a first-party Foundry service. In Microsoft's words in the blueprint, "Foundry provides the Azure-native platform for deployment, governance, and billing; Fireworks is the inference layer serving the open models behind them."

According to Unite.AI, the catalog's 26 models included Moonshot AI's Kimi K2.5, DeepSeek V3.2, MiniMax M2.5, OpenAI's gpt-oss-120b and a DeepSeek V4 Pro described as a 1.6-trillion-parameter flagship. Six of the 26, including Kimi K2.6 and Z.ai's GLM-5.1, offered pay-per-token billing at the time; the other 20 ran only on provisioned throughput units (PTUs), which is reserved capacity billed by the hour. Microsoft keeps the current list in its catalog reference, and Fireworks says models are added on a rolling basis, so the count can change.

Diagram of Microsoft's startup blueprint: an app on Azure Container Apps calls API Management, then a Fireworks endpoint deployed via Foundry, with Container Registry, Key Vault, Redis and Monitor

What startups build

The blueprint runs inside the startup's own Azure subscription. An app hosted on Azure Container Apps sends requests to a Fireworks model endpoint deployed through Foundry. Azure Container Registry stores the app's container images, Azure Key Vault holds credentials and endpoint details, and Azure Monitor (optional) tracks performance.

Microsoft describes a path that starts small and grows:

  • Deploy a single model, route traffic through API Management, and track latency, usage and cost.
  • Add Azure Cache for Redis to avoid paying twice for repeated requests.
  • Run multiple model variants for A/B testing as traffic grows.
  • Fine-tune a model with Fireworks Training and bring the weights back to Azure through a bring-your-own-weights workflow.

Microsoft's pitch is about cost control. The blueprint calls inference "one of the largest controllable cost drivers for AI-native companies" and tells teams to "track cost per million tokens as a core engineering metric." It also argues that serving open models this way lets a startup "choose, optimize, and switch the models behind your product" without being tied to one model provider.

The fine print on credits and compliance

The credits do not cover everything. Microsoft's blueprint says Startup credits apply to Fireworks deployments using Data Zone Standard, a pay-per-token option, and that PTUs "are reserved capacity and not covered by Startup credits." Since most of the catalog ran only on PTUs, according to Unite.AI's count, credits reach a smaller set of models than the headline number suggests.

The compliance limits matter for regulated buyers. Microsoft's setup documentation states that Fireworks on Foundry:

  • is currently excluded from Microsoft's EU Data Boundary commitments;
  • has not achieved FedRAMP, the US government cloud security authorization;
  • should not be used to store, process or transmit payment card data.

Microsoft's privacy FAQ adds that customer data "travels outside Microsoft facilities because Fireworks runs inferencing for this service on Fireworks' GPUs," though processing stays governed by Microsoft's terms because Fireworks acts as a Microsoft subprocessor. The Data Zone Standard pay-per-token option is limited to US regions: Fireworks' documentation lists six, while Microsoft's catalog reference now lists seven. Microsoft's catalog reference also lists a Global Standard pay-per-token option, available in all global Azure regions except Azure Government. Fireworks' documentation sets the throughput limit for its pay-per-token mode at 500,000 tokens per minute.

The same setup page carries a transparency note: Microsoft "doesn't develop, train, fine-tune, or evaluate the safety, security, or Responsible AI characteristics" of these models, and customers are "solely responsible for assessing the suitability of any model" for their intended use.

Billing terms also shift. Unite.AI reported that pay-per-token billing for GLM-5.1 and MiniMax M2.5 was deprecated effective August 7, 2026, with both still available on provisioned throughput, and that pay-per-token offers for four other catalog models were already deprecated.

Why it matters

For startups, credits they already hold can pay for pay-per-token experiments, and switching between models does not mean switching clouds. Unite.AI calls the blueprint Microsoft's bid to make Fireworks on Foundry "the default starting stack for AI-native companies building on Azure."

Unite.AI also places it in a longer build-out of Foundry's model catalog. Moonshot AI was later reported to be in revenue-sharing talks with Microsoft, Amazon and Google to host Kimi K3, and Azure's AI customer list has grown to include Meta as a major buyer. For a plain-English guide to the models themselves, see our list of the best open-weight models as of October 2026.

The catalog has already moved since August. As of October 27, 2026, Microsoft's catalog reference lists pay-per-token deployments for newer models such as Kimi K3, GLM-5.3 and DeepSeek V4.1 Flash, and shows the pay-per-token offerings for models from the August list such as Kimi K2.5, DeepSeek V3.2 and gpt-oss-120b as retired. What to watch: whether credit eligibility widens to PTUs and whether the compliance gaps close.

Sources

NewsMicrosoftAzureOpen WeightsFireworks AI