Azure models beyond OpenAI

Beyond Azure OpenAI (GPT, o-series, embeddings, etc.), Microsoft Foundry and your Azure AI Services account can give you access to models from other providers. Availability depends on region and how the resource is set up.

Reference: Foundry models sold directly by Azure (switch to “Other model collections” for non-OpenAI).


Other providers (via Azure)

ProviderExample modelsUse case
DeepSeekDeepSeek-V3-0324, DeepSeek-V3.1, DeepSeek-V3.2, DeepSeek-R1, DeepSeek-R1-0528Chat, reasoning, coding. Deploy on brainforge-openai-eastus2 (East US was legacy; migrate any old deployments).
CohereCohere-command-a, embed-v-4-0, Cohere-rerank-v4.0-pro, Cohere-rerank-v4.0-fastChat, embeddings, rerank.
MetaLlama-4-Maverick-17B-128E-Instruct-FP8, Llama-3.3-70B-InstructOpen-weight LLMs.
Mistralmistral-document-ai-2505, Mistral-Large-3Chat, document AI.
xAIgrok-3, grok-3-mini, grok-4, grok-4-fast-reasoning, grok-4-fast-non-reasoning, grok-code-fast-1Chat, reasoning, code.
MicrosoftMAI-DS-R1, model-routerMicrosoft models, routing.
Moonshot AIKimi-K2-Thinking, Kimi-K2.5Reasoning, long context.
Black Forest LabsFLUX.2-pro, FLUX.1-Kontext-pro, FLUX-1.1-proImage generation.

What you already have (brainforge-openai-eastus2)

Policy: New deployments only on East US 2. brainforge-openai (East US) is legacy/deprecating.

From az cognitiveservices account deployment list -g brainforge -n brainforge-openai-eastus2, typical deployments include OpenAI chat, embeddings, and (when deployed) third-party models on the same resource. See azure-openai-deployment-options.md and azure-models-for-devs.md.


How to use non-OpenAI models

  • Same resource (brainforge-openai-eastus2): Non-OpenAI deployments (e.g. DeepSeek) use the same Azure OpenAI–compatible endpoint and key as GPT deployments; the deployment name selects the model. Call via deployments/{deployment-name}/... on the East US 2 base URL.
  • New resource: For providers that are only in Foundry (e.g. some Cohere, Mistral, xAI), you may need a Foundry project or a different resource type; see Foundry models and the AI Model Catalog.
  • Region: Deploy on East US 2 only for new work; check each model’s region table in Azure docs.

Kimi (Moonshot) K2-Thinking & K2.5 – optional for devs

The platform default is the best OpenAI model we have (azureModel_default, upgradeable via AZURE_DEFAULT_CHAT_*). Kimi is opt-in only: agents use it only when you explicitly set model: azureModel_kimi_k2_thinking or azureModel_kimi_k2_5.

After deploying Kimi-K2-Thinking and/or Kimi-K2.5 in Azure (same resource or Foundry), set in .env:

  • AZURE_KIMI_K2_THINKING_BASE_URL (optional; defaults to Azure default chat / o4-mini / gpt-4o base URL – never non-Azure OPENAI_BASE_URL)
  • AZURE_KIMI_K2_THINKING_DEPLOYMENT (default: Kimi-K2-Thinking)
  • AZURE_KIMI_K2_5_BASE_URL (optional; same Azure-only fallback)
  • AZURE_KIMI_K2_5_DEPLOYMENT (default: Kimi-K2.5)

In code, use the exported models so agents can opt in:

import { azureModel_kimi_k2_thinking, azureModel_kimi_k2_5 } from "@/mastra/azure";
 
// Use K2-Thinking for an agent
model: azureModel_kimi_k2_thinking
 
// Or K2.5
model: azureModel_kimi_k2_5

If Kimi env is not set, these exports fall back to o4-mini so the app keeps running without a Kimi deployment.


Quick reference