Fireworks on Foundry

Fireworks on Microsoft Foundry brings Fireworks’ best-in-class inference for open-source models into Foundry as a first-party service. Access state-of-the-art open models through a single Azure endpoint, deploy your own custom weights, and scale in production with Provisioned throughput, all with Azure governance, billing, and controls built in. It’s generally available as of June 2, 2026.

Fireworks on Foundry offers day-zero access to leading open models through serverless pay-as-you-go and Provisioned throughput. Launched models include OpenAI gpt-oss-120b, Kimi K series, DeepSeek, MiniMax and more. New model families are added regularly. Browse the full list in the Foundry model catalog under the Fireworks collection.

Yes. With bring-your-own-weights (BYOW), you can deploy custom or fine-tuned open models (full-weight, with LoRA coming soon) on Provisioned throughput as long as the base model architecture is supported. Fine-tune with your preferred tools, then deploy at scale on Fireworks’ optimized inference backend through Foundry. Custom-weight deployments are priced the same as the underlying base model.

Choose the option that fits your workload. Serverless pay-as-you-go bills per token is ideal for experimentation and variable traffic. Provisioned Throughput Units (PTUs) deliver predictable, steady-state performance with latency and availability SLAs, and can be applied flexibly across your Foundry model deployments. Consumption is billed directly through Azure, no separate marketplace contract. See the Foundry pricing page for current rates.

Yes. Fireworks on Foundry is delivered as a first-party Foundry service, so you get Azure governance, access controls, observability, and billing, plus private VNet networking. Models work with the Foundry Agent service and tools like Azure AI Search and Fabric Data Agent. Note: during onboarding, deployment requires a per-subscription opt-in, and it shouldn’t be used for workloads requiring FedRAMP or PCI compliance.