Skip to content

Model catalog: Azure cloud

Scope: Azure AI Foundry

This page describes the Azure AI Foundry target, the hosted-cloud target of ADR-0011. Foundry Local and Azure Local Foundry differ from it in models, features, identity, cost, and operations. Compare all three on Deployment targets.

This is a curated shortlist, not a menu

This page records what this project chose, considered, and rejected. It is not a list of what Azure AI Foundry can run. It holds roughly 38 rows; a single Azure region offers several hundred models. If you want to know what is available rather than what was chosen, go to Available models: Azure AI Foundry. The distinction is explained in chosen versus available.

This catalog is the living, human-readable list of the models this repo's Azure AI Foundry methodology has deployed, evaluated, or rejected on the Azure cloud target. It is the single source of truth for "which models has this project put on the table, and where does each one stand" for that target.

It began as the model set this project originally needed, and it has grown by evaluation since. A model absent from it is not unavailable; it is unevaluated, or it was evaluated and rejected, in which case there is a row below saying so. A reader adopting this methodology rather than this project should expect their own catalog to differ: the methodology transfers, the roster does not.

None of these models runs on either on-premises target: those run a separate open-weight roster with no image generation and no text to speech, listed in full in Available models: Foundry Local and Azure Local Foundry. See the reference index for why the lists are split, and the deployment targets comparison for what each target can run. The machine-readable counterpart is the model registry (models/registry.schema.json, with models/registry.starter.json as the real starting roster), which a consuming project resolves at runtime; this catalog is the prose record a human reads to understand the why behind each row.

Adding a new model to the evaluation set is a catalog row plus a registry entry, not a new ADR. The general selection methodology lives in one place, ADR-0002 for image models and ADR-0003 for voice, and each new candidate is evaluated against that methodology and recorded here. A model gets its own ADR only when adopting it changes a decision the methodology cannot already express (for example a new region, a new resource, or a new access-governance posture). The FLUX adoption is the worked example: it was originally captured as its own record and is now folded back into the general image-selection ADR plus these catalog rows, so a future image model does not need a new ADR to be added.

Status vocabulary

StatusMeaningRegistry mapping
deployedLive on a Foundry resource and callable by the pipeline.status: deployed
availableIn the catalog and deployable in-region with no blocker, but not adopted (a candidate held for a future A-B arm or backfill).status: planned
evaluatedResearched against the methodology and passed over for the current workload, but not a hard reject (region, access gate, or style bias made it wrong for now; revisit if the constraint changes).status: planned or omitted
rejectedRuled out and kept on the record so the decision is not re-researched later (retired, off-catalog, or structurally unfit).status: rejected (never deleted)

The registry enum is deliberately narrower (deployed / planned / rejected); this catalog splits the middle into available and evaluated for human clarity. Never delete a rejected row: the point is to record why a candidate was passed over so it is not re-litigated.

Image models

Selection methodology and access model: ADR-0002. Historical FLUX adoption record: ADR-0010 (superseded). Research behind these rows: SPIKE-01 (MAI image family) and SPIKE-12 (broader catalog and video alternatives).

ModelProviderStatusRegion(s)Key traitsNotes / why or why notSource
MAI-Image-2.5Microsoftdeployed (baseline)East US (also West Central US, West US, West Europe, Sweden Central, South India, UAE North)Preview. Text-to-image plus image-to-image edits. Thin documented API: model, prompt, image (edits), width, height only, so no seed, mask, negative prompt, style-reference, or candidate-count. One image per call. Canvas min 768x768, max total 1,048,576 px.First-party in-region baseline and A-B comparison point. Driven text-only it drifted on a consistent non-photorealistic house style and on character identity (no reference-image conditioning on the generation path), which is exactly what drove adding the FLUX roster. Retained as fallback, not removed.MAI image how-to, Models sold by Azure
MAI-Image-2.5-FlashMicrosoftdeployed (draft arm)Same as MAI-Image-2.5Preview. Same two endpoints and same parameter set as 2.5; positioned as faster and lower cost per token. Same RPM ceiling as 2.5 (buys latency and cost, not throughput).Deployed as the cheap draft and thumbnail arm. A cost lever rather than a capability gain, which is why it sits alongside 2.5 rather than replacing it.MAI image how-to
MAI-Image-2.5-ProMicrosoftdeployed (high fidelity)East US (per region-availability matrix)Preview. Positioned for more photo-realistic imagery than MAI-Image-2.5.Deployed as the highest-fidelity arm of the MAI family. Its photo-realistic bias makes it the wrong default for a stylised house look, so it is selected per call rather than used as the baseline.Models sold by Azure
FLUX.2-proBlackForestLabsdeployed (primary)East US (all Americas regions, Global Standard)GA, no retirement date. Multi-reference conditioning up to 8 reference images. Output up to 4 MP. Highest fidelity of the FLUX family.Primary scene-art generator. Supplies the multi-reference character-and-style lock the MAI text-only surface structurally lacked (feed locked character sheets plus a style anchor into each generation). Adopted in ADR-0010, now carried by ADR-0002 plus this row.Deploy FLUX models, Models sold by Azure
FLUX.1-Kontext-proBlackForestLabsdeployed (single-reference companion)East US (all Americas regions, Global Standard)GA. In-context editing anchored to one reference image (text and 1 image, 1 MP). Documented for character consistency and advanced editing.Single-reference companion: hold one identity or look stable across an edit where composing from many references is not the point.Deploy FLUX models, Models sold by Azure
FLUX-1.1-proBlackForestLabsdeployed (fast exploration)East US (all Americas regions, Global Standard)GA. Text-to-image, fast inference, strong prompt adherence, competitive pricing. No multi-reference lock.Fast, cheap exploration and low-stakes passes, not locked-character final frames.Deploy FLUX models, Models sold by Azure
FLUX.2-flexBlackForestLabsavailableEast US (all Americas regions, Global Standard)GA. Text plus up to 10 reference images (2 more than FLUX.2-pro). Exposes guidance and steps for fine control. Output up to 4 MP. Positioned for text-heavy layouts and text overlay.Deployable in-region today, but an incremental extension of the same FLUX approach already carrying the style-lock load, not a new capability class. Its text-overlay strength is marginal for wordless scene art where captions are composited later. Would be added as an A-B arm behind its own ADR if wanted.Deploy FLUX models, Region availability
GPT-image-2OpenAI (Azure OpenAI)evaluatedEast US 2 and West US 3 only in the Americas (NOT East US)GA, no access application. Mask-based inpainting and variations, arbitrary resolutions up to 4K (long edge up to 3,840 px, aspect up to 3:1), 1 to 10 images per request via n. Realism-optimized with advanced face preservation.Passed over for the non-photorealistic house style: realism and face-preservation bias is the wrong direction, and it is out of region for an East US account (would need a second resource in eastus2 or westus3). Its unique capability (mask inpainting) is not required by the prompt-plus-multi-reference pipeline. Revisit only if a masked-edit workflow becomes a real need.Image models and capabilities, Region availability
gpt-image-1.5OpenAI (Azure OpenAI)evaluatedEast US 2 and West US 3 only in the AmericasLimited-access preview (apply at aka.ms/oai/gptimage1.5access). Realism-optimized, face preservation, improved speed and cost over gpt-image-1. Inpainting and variations with mask plus prompt.Same realism bias and same out-of-region problem as GPT-image-2, plus a limited-access application gate. Not fit for the non-photorealistic style.Image models and capabilities
gpt-image-1OpenAI (Azure OpenAI)evaluatedEast US 2 and West US 3 only in the AmericasLimited-access preview (apply at aka.ms/oai/gptimage1access). Realism-optimized, face preservation. Inpainting and variations with mask plus prompt.Older GPT-image variant. Same realism bias, out-of-region, and access-gate issues; no advantage over the newer variants for this workload.Image models and capabilities
gpt-image-1-miniOpenAI (Azure OpenAI)evaluatedEast US 2 and West US 3 only in the AmericasLimited-access preview (apply at aka.ms/oai/gptimage1access). Cost-efficient and faster for bulk or iterative generation. No dedicated face preservation, better for non-portrait general creative imagery.Cheapest GPT-image tier, but still out of region for an East US account and behind an access application; the FLUX roster already covers cheap exploration in-region.Image models and capabilities
DALL-E 3 (dall-e-3)OpenAI (Azure OpenAI)rejectedRetired (previously East US, Australia East, Sweden Central)Retired on 2026-03-04. No longer available for new deployments; existing deployments are non-functional. Replaced by the GPT-image series.Not selectable. Kept on the record so it is not re-proposed. Its official replacement path (GPT-image) is the evaluated set above, which is itself out of region and realism-biased for this workload.Image generation how-to (retirement note)
Stable Diffusion (Stability AI: Stable Diffusion 3.5 Large, Stable Image Core, Stable Image Ultra)Stability AIrejectedFoundry model catalog serverless (partner) deployment; not in "Foundry Models sold by Azure"Available in the broader Foundry model catalog as a Stability AI serverless (pay-as-you-go) deployment, NOT as a first-party sold-by-Azure model. Realism and photorealism tuned (Stable Image Ultra is positioned for photorealism and product imagery).Rejected for this build, not because it is absent from Azure (it is present as a partner serverless model) but because it is off the shared sold-by-Azure AIServices resource this methodology deploys onto (a separate serverless endpoint and governance path) and carries a realism bias, not the non-photorealistic illustrated style required. Revisit only if a separate serverless deployment is justified.Stability AI in Foundry catalog, Deploy Stability AI models
ImagenGooglerejectedNot offered in the Azure AI Foundry catalog (Google Cloud Vertex AI only)Google's image family (Imagen) is a Google Cloud Vertex AI offering. It was not found in the Azure AI Foundry model catalog during research: the first-party image families sold by Azure are only MAI-Image, FLUX, and GPT-image.Rejected as an off-Azure departure: adopting it would leave the Azure AI Foundry catalog, the shared resource, and this repo's Azure-first, harness-only governance posture. Not deployable on the shared resource.Models sold by Azure (image families)

Cost note for image models (applies to every deployed and available row)

All Foundry image models bill on the token-metering model (per token in and out, no charge for the resource or the deployment itself). The exact per-token rate for the MAI and FLUX image models could not be verified from a rendered first-party pricing page during SPIKE-01 and SPIKE-12, and tokens-per-generated-image is not published by Microsoft for any MAI image model. Treat per-image cost as UNKNOWN until measured: read the live rate in the Foundry portal at deploy time and run a two-call cost probe before any batch. See SPIKE-01, Q2 and ADR-0006 for the budget guardrail.

Voice models

Selection methodology and locked voice set: ADR-0003. Research: SPIKE-02 and the speech-model survey in SPIKE-07.

ModelProviderStatusRegion(s)Key traitsNotes / why or why notSource
MAI-Voice-2Microsoft (MAI)deployed (expressive baseline)East USExpressive prebuilt voices plus mstts:express-as styles; same Speech SDK/endpoint, no Foundry deployListen-only (no documented WordBoundary); 22 USD per 1M characters; Preview. Reached by SSML voice name, so it is not a deployment resource and its registry id sits in skipDeploymentModelIds.https://learn.microsoft.com/azure/ai-services/speech-service/mai-voices
azure-neural-standardMicrosoftdeployed (timing baseline)East USStandard neural voices and Dragon HD Omni on the same Speech resourceThe counterpart to MAI-Voice-2, not a replacement. These voices do emit WordBoundary events and viseme identifiers, which the expressive preview model cannot, so they carry any read-along or lip-sync track. Also SSML-selected, so also not a deployment resource.https://learn.microsoft.com/azure/ai-services/speech-service/language-support
Azure neural / Dragon HD Omni (native)Microsoft (Azure Speech)availableEast US (per-region matrix)Emits WordBoundary; native visemes (redlips_front en-US, blend shapes en-US/zh-CN)The read-along and lip-sync baseline; standard neural documented near 15 USD per 1M charactershttps://learn.microsoft.com/azure/ai-services/speech-service/high-definition-voices
ElevenLabsElevenLabsevaluatedExternal SaaS (off-Azure)Character-level timestamps (word derivable) plus a hosted Forced Alignment endpoint; no phoneme/visemeDoes not beat the in-resource baseline for word-sync; separate billing and egresshttps://elevenlabs.io/docs/api-reference/text-to-speech/convert-with-timestamps
Cartesia (Sonic)CartesiaevaluatedExternal SaaSNative word plus phoneme timestamps (add_timestamps / add_phoneme_timestamps), streamingOne of only two non-Azure vendors emitting native phoneme timing; lip-sync candidatehttps://docs.cartesia.ai/api-reference/tts/websocket
Hume Octave 2Hume AIevaluatedExternal SaaSNative word plus phoneme (IPA) timestamps; emotional delivery controlNative phoneme track for lip-sync; external SaaShttps://dev.hume.ai/docs/text-to-speech-tts/timestamps
RimeRimeevaluatedExternal SaaSWord timestamps on the WebSocket path (HTTP path lacks them)Word-sync only, integrator-documented; confirm first-party before adoptionhttps://docs.livekit.io/agents/models/tts/rime/
Deepgram Aura / Aura-2DeepgramevaluatedExternal SaaSStreaming TTS with token-by-token inputTTS word timestamps UNKNOWN (word timing is a Deepgram STT feature); not confirmed for read-alonghttps://deepgram.com/learn/aura-text-to-speech-adds-websocket-support-for-input-streaming
OpenAI TTS (gpt-4o-mini-tts, tts-1)OpenAIrejectedExternal SaaS (also via Azure OpenAI)Delivery control via instructions; no timestamp outputTimestamps are STT-only; fails the word-sync barhttps://developers.openai.com/api/docs/guides/text-to-speech
PlayHTPlayHTevaluatedExternal SaaSStreaming TTSNative word/phoneme timestamps UNKNOWN (no first-party timestamp doc located); capability unverifiedhttps://play.ht/
Kokoro-82Mhexgrad (open weights)evaluatedSelf-host (CPU-capable)Apache 2.0; phoneme-based (misaki G2P, IPA); 82M paramsCommercial-safe open option; no native word-timestamp API (word timing via forced alignment)https://huggingface.co/hexgrad/Kokoro-82M
PiperRhasspy (open weights)evaluatedSelf-host (CPU real-time)MIT; eSpeak NG phoneme-based; fastest on CPUNo native word timestamps (pair with forced alignment); dev moved to OHF-Voice/piper1-gpl (GPL)https://github.com/rhasspy/piper
ChatterboxResemble AI (open weights)evaluatedSelf-host (GPU preferred)MIT; ~0.5B Llama backbone; emotion control, zero-shot cloningCommercial-safe; native timing not documented (forced alignment applies)https://www.resemble.ai/learn/models/chatterbox
Coqui XTTS v2Coqui (open weights)rejectedSelf-hostCPML non-commercial license; Coqui wound down early 2024Excluded on license for commercial publishinghttps://huggingface.co/coqui/XTTS-v2
F5-TTSSWivid (open weights)rejectedSelf-hostCC-BY-NC-4.0 weights (non-commercial)Excluded on license for commercial publishinghttps://huggingface.co/SWivid/F5-TTS

Video models

Research: SPIKE-12 (Sora, Sora 2, and non-Azure image-to-video vendors). The core video path is locked graphite stills plus Ken Burns pan and zoom plus narration via ffmpeg; Sora and Sora 2 are rejected for the core illustrated look (realism-tuned, eastus2-only, and Sora 2 rejects image-to-video inputs containing human faces).

ModelProviderStatusRegion(s)Key traitsNotes / why or why notSource
SoraOpenAI (sold by Azure)rejectedeastus2 only (Americas)Text-to-video / image-to-video; realism and physics tunedRejected for the core illustrated look: realism bias will not hold a stylized look; separate resource from an East US accounthttps://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
Sora 2OpenAI (sold by Azure)rejectedeastus2 only (Americas)text/image/video-to-video, output audio, remix; realism/physics/temporal tuned; PreviewRAI blocks IP/photorealistic and rejects input images with human faces (blocks image-to-video for people-present art); eastus2-onlyhttps://learn.microsoft.com/azure/foundry/openai/concepts/video-generation#responsible-ai-and-video-generation
Off-Azure (Runway Gen-4/4.5, Luma, Kling, Pika)Runway / Luma / Kling / PikarejectedOff-AzurePreserve a stylized source look better than Sora (character consistency, camera control)Governance departure: off the harness, off Azure, separate RAI and billing, no HCS Governance MCP coverage; capability claims are roundup-grade, not first-partyhttps://runwayml.com/

Reasoning / review models

Research: SPIKE-10 (latest GPT), SPIKE-11 (newer Grok), and SPIKE-15 (niche reviewers).

Fourteen reasoning models are deployed. The roster was deliberately widened from the original two-model reviewer pair to the full family where a vendor offers one (flagship plus siblings), because these are pay-as-you-go base models: an idle deployment costs nothing, so the cost of having a model available is zero and the cost of not having it is a redeploy at the moment you need it.

Note the split on vision. Only the grok-4.1-fast pair is vision-capable; every newer-numbered Grok is text-only. So the newest model is not automatically the right one, and a reviewer that must actually look at a generated image has exactly one Grok option.

ModelProviderStatusRegion(s)Key traitsNotes / why or why notSource
gpt-5.6-solOpenAI (sold by Azure)deployed (flagship)East US (Global Standard)Vision; reasoning; GAFamily flagship. Primary reviewer where quality matters more than cost.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
gpt-5.6-terraOpenAI (sold by Azure)deployed (balanced default)East US (Global Standard)Vision; reasoning; 1.05M-token context; GABalanced sibling with the largest context of the family. Default for document and image review.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
gpt-5.6-lunaOpenAI (sold by Azure)deployed (economy)East US (Global Standard)Vision; GACheapest sibling, for high-volume or low-stakes passes.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
grok-4-1-fast-reasoningxAI (sold by Azure)deployed (only vision-capable Grok)East US (Global Standard)Vision; reasoning; 128K in / 128K out; GAThe only Grok that can grade an image. Required for any Grok reviewer in the image path.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
grok-4-1-fast-non-reasoningxAI (sold by Azure)deployedEast US (Global Standard)Vision; 128K; GALower-latency counterpart for vision tasks that do not need a reasoning pass.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
grok-4.3xAI (sold by Azure)deployed (newest Grok)East USText-only; Preview; 200K in / 8K outNewest Grok, deployed for text review. Text-only, so it cannot serve the image-review role.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
grok-4-20-reasoningxAI (sold by Azure)deployedEast USText-only; Preview; 262K in / 8K outLargest Grok context window. Text-only, same constraint as above.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
grok-4-20-non-reasoningxAI (sold by Azure)deployedEast USText-only; Preview; 262K in / 8K outNon-reasoning counterpart of the same generation, for lower-latency text passes.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
Kimi-K2.7-CodeMoonshot AI (sold by Azure)deployed (code reviewer)East US (Global Standard)Coding-specialized; text+image 262K; tool calling; PreviewThe code and document reviewer, distinct from the image-review role.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
Kimi K3 (2026-07-16 flagship)Moonshot AIevaluated (watch item)Not in Foundry catalog~2.8T MoE, 1M-token context (third-party reported)NOT in the Azure catalog yet; track for a future Kimi-K3 entry.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
DeepSeek-V4-ProDeepSeek (sold by Azure)deployedEast US (Global Standard)Strong code/reasoning, text-focusedIndependent second opinion on code and structured reasoning, where depth matters.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
DeepSeek-V4-FlashDeepSeek (sold by Azure)deployedEast US (Global Standard)Faster, cheaper sibling of V4-ProThe bulk-pass arm of the same family.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
Llama-4-Maverick-17B-128EMeta (sold by Azure)deployedEast US (Global Standard)Text+image up to 1M-token context; general reviewerLong-document comprehension. Note its catalog name is far longer than the deployment name, which the catalog generator resolves by prefix.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
Mistral-Large-3Mistral AI (sold by Azure)deployedEast US (Global Standard)Credible general reviewerVendor diversity, useful where model provenance matters.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
Phi-4-reasoningMicrosoft (sold by Azure)deployedEast US (Global Standard)Small, cheap, first-party reasoningLow-cost pre-filter that escalates only nontrivial findings. Also the closest match to what the on-premises track can run.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
Phi-4-mini-reasoningMicrosoft (sold by Azure)evaluatedEast US (Global Standard)Smaller Phi variantNot deployed; the full Phi-4-reasoning is already cheap enough.https://learn.microsoft.com/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure-region-availability
FW-GLM-4.x, FW-Qwen3.6, FW-MiniMax-M2.5, FW-GPT-OSS-120Bvia FireworksevaluatedEast US (Fireworks partner path)Strong coding/agentic optionsHeavier governance caveats (Fireworks data-sharing, no EU Data Boundary/FedRAMP); prefer the sold-by-Azure pathhttps://learn.microsoft.com/azure/foundry/how-to/fireworks/enable-fireworks-models

See also

  • Machine-readable registry: models/registry.schema.json, models/registry.starter.json, and the consumption contract in the model registry guide.
  • Image selection methodology and access: ADR-0002.
  • Historical FLUX adoption record (superseded): ADR-0010.