Microsoft has launched 26 open models accessible to startups through its Fireworks AI integration on Foundry. This initiative, announced on August 4, 2026, aims to equip AI-native companies with tools to optimize their development processes. Startups in the Microsoft for Startups program can utilize their Azure credits, potentially amounting to $150,000, for Fireworks model deployments. The Fireworks system can now be leveraged for a range of open-weight models from suppliers such as DeepSeek, Moonshot AI, and OpenAI’s gpt-oss line, all housed within Azure’s model catalog and governed by Azure’s management framework.
The reference architecture designed for these startups allows them to operate entirely within their own Azure subscriptions. It employs a containerized application on Azure Container Apps that calls Fireworks model endpoints via Foundry. This arrangement ensures that essential components, such as Azure Container Registry for image storage and Azure Key Vault for credential management, are integrated seamlessly. The framework scales by routing traffic through Azure API Management and utilizes Azure Cache for Redis to minimize redundant inference calls, providing scalability in response to demand. Notably, monitoring tools within Azure also enable teams to track latency, error rates, and token consumption effectively.
Among the 26 models available, notable entries include Moonshot AI’s Kimi K2.5 and DeepSeek V4 Pro, which features a substantial 1.6-trillion-parameter architecture. While six models offer pay-per-token billing, the remainder operate on Azure’s reserved-capacity pricing. The architecture also supports importing custom or fine-tuned weights via a bring-your-own-weights approach, with LoRA adapter compatibility in public preview.