Microsoft MAI and Phi Models: Every Model, Spec, Price and Release Date

Microsoft publishes two model families of its own: MAI, the proprietary models from Microsoft AI, whose seven-model family was announced at Microsoft Build on 2 June 2026, and Phi, a line of small open-weight language models released under the MIT licence on Hugging Face. This page lists every current MAI and Phi model named on Microsoft's own pages, with release dates, parameters, context windows, licences, published prices, availability and model card links. MAI facts come from microsoft.ai model pages, blog posts and the Build 2026 keynote transcript; Phi facts come from Microsoft's Hugging Face model cards. Where Microsoft publishes no value, the page says "Not stated by Microsoft" rather than estimating. Facts are as of 15 November 2026.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• MAI Thinking 1 is Microsoft's own reasoning model. A sparse Mixture-of-Experts model with 35B active and about 1T total parameters and a 256k context window, in public preview on Microsoft Foundry, according to Microsoft AI's introduction post of 12 August 2026.
• MAI covers text, code, image, speech and security. The microsoft.ai models page lists MAI Thinking, MAI Code, MAI Cyber, MAI Image, MAI Transcribe and MAI Voice, plus custom models through Microsoft Frontier Tuning.
• Phi models are open weight under MIT. Phi-4 (14B), Phi-4-mini-instruct (3.8B), Phi-4-multimodal-instruct (5.6B), Phi-4-reasoning (14B) and Phi-4-reasoning-vision-15B all carry the MIT licence on their model cards.
• MAI Cyber 1.1 Flash is restricted. Microsoft states that verified defenders can access it through Codename MDASH, with a register-interest route.
• Few MAI token prices are published. Microsoft lists prices for MAI Transcribe 2 and MAI Voice 2.1 on their model pages; per-token prices for MAI Thinking 1 and MAI Code are not stated on the pages checked.

What is Microsoft MAI and Phi?

Microsoft AI, the division behind Copilot, develops the MAI models. The MAI Thinking 1 post attributes the model to the Microsoft AI Superintelligence team. MAI models reach users in three ways: inside Microsoft products (Copilot, GitHub Copilot, VS Code, PowerPoint, Teams), through Microsoft Foundry for developers, and through the MAI Playground, which Microsoft describes as intended for personal, non-commercial use only. The Build 2026 keynote also named third-party distribution for some MAI models, including OpenRouter, Fireworks and Baseten.

Phi is a separate research line of small language models published as open weights on Hugging Face under the microsoft organisation, from phi-1 (2023) to the Phi-4 family. Phi model cards describe target use in memory or compute constrained environments and latency-sensitive applications. MAI and Phi are listed separately throughout this page.

MAI names follow the pattern MAI <capability> <version> with an optional Flash suffix for the faster, lower-cost variant (for example MAI Voice 2.1 Flash). Phi names encode generation, size and specialisation (Phi-4-mini-reasoning, Phi-4-reasoning-vision-15B). See LLM Model Naming and Versioning Explained.

Current MAI and Phi models at a glance

As of 15 November 2026. Sources: microsoft.ai model pages and Microsoft Hugging Face model cards. MAI models in the order of the microsoft.ai models page, then Phi models newest first.

For Phi models the Hugging Face repository name is the identifier. The microsoft.ai pages for MAI models give no API model strings.

Key facts for the main MAI and Phi models

As of 15 November 2026. Identical fields for every model; "Not stated by Microsoft" means not stated by Microsoft on the pages checked (2026-11-15).

Parameter labels: several Hugging Face pages show a rounded size in the page header that differs from the model card text (for example 15B in the header and 14B in the Phi-4 card). This page uses the model card text. MAI Thinking 1 uses a sparse Mixture-of-Experts design; see Mixture of Experts Explained.

MAI and Phi model families and naming

• MAI Thinking: MAI Thinking 1, Microsoft's first MAI text reasoning model, trained without distillation from third-party models according to its introduction post. Compatible with the Chat Completions API.
• MAI Code: MAI Code 1 Flash (5B, announced at Build as a default model for about 10% of individual VS Code users initially, and in GitHub Copilot CLI), followed by MAI Code 1.1 Flash, which can also run locally in GitHub Copilot on supported RTX Spark devices.
• MAI Cyber: MAI Cyber 1.1 Flash, built into Codename MDASH, a multi-agent harness for finding and fixing vulnerabilities.
• MAI Image: MAI Image 2.5 and 2.5 Flash (Build), then MAI Image 2.6 and 2.6 Flash.
• MAI Transcribe: MAI Transcribe 1.5 (Build, 43 languages), then MAI Transcribe 2 and MAI Transcribe 2 Streaming (60 languages).
• MAI Voice: MAI Voice 2 (15 languages) and MAI Voice 2 Flash (Build), then MAI Voice 2.1 and 2.1 Flash (23 languages).
• Phi-4 family: phi-4, Phi-4-mini-instruct, Phi-4-multimodal-instruct, Phi-4-reasoning, Phi-4-reasoning-plus, Phi-4-mini-reasoning, Phi-4-mini-flash-reasoning and Phi-4-reasoning-vision-15B, with ONNX and GGUF builds of several.
• Other Phi models: Phi-mini-MoE-instruct, Phi-tiny-MoE-instruct, Phi-Ground and Phi-Ground-Any, and the earlier Phi-3.5, Phi-3, phi-2, phi-1.5 and phi-1 repositories.

When was each MAI and Phi model released?

As of 15 November 2026, in date order. Phi dates marked "repository" are the Hugging Face repository creation dates from the Hugging Face API; other dates come from model cards, posts or the Build keynote.

Release dates for MAI Code 1.1 Flash, MAI Cyber 1.1 Flash, MAI Image 2.6, MAI Transcribe 2 and MAI Voice 2.1 are not stated on their microsoft.ai pages, which were all last modified on 10 or 11 November 2026.

How much do MAI and Phi models cost?

As of 15 November 2026. Sources: microsoft.ai model pages. Prices are in USD as listed.

For how token, cached and batch pricing work in general, see LLM Pricing Explained: Tokens and Caching. To cap spend on keys, see How to set spend limits on LLM API keys.

Where are MAI and Phi models available?

• Microsoft Foundry: MAI Thinking 1 in public preview; MAI Image 2.5 and MAI Transcribe 1.5 named for Foundry in the Build keynote.
• Microsoft products: MAI Code in GitHub Copilot and VS Code; MAI Transcribe 1.5 in Copilot, Teams, GitHub and Dynamics 365 Contact Centre; MAI Image 2.5 in PowerPoint and rolling out to OneDrive (Build keynote).
• MAI Playground: MAI Thinking 1, Image 2.6, Transcribe 2 and Voice 2.1 can be tried after Microsoft sign-in; Microsoft states it is for personal, non-commercial use only.
• Third-party platforms (Build keynote): OpenRouter for MAI Image 2.5, MAI Transcribe 1.5 and MAI Voice 2; Baseten for MAI Thinking 1; Baseten, Fireworks, Foundry and OpenRouter planned for MAI Code 1 Flash.
• Codename MDASH: MAI Cyber 1.1 Flash, for verified defenders.
• Phi open weights: Hugging Face under the microsoft organisation, with ONNX builds of several models and a GGUF build of phi-4 for local runtimes.

Running Phi weights yourself moves security responsibility to you. See Running LLMs Locally Explained and the Open-Weight Model Security Guide.

Restricted-access MAI and Phi models

As of 2026-11-15.

What is the context window of MAI and Phi models?

MAI Thinking 1 has a 256k token context window, stated in both the Build keynote and its introduction post. Context windows for MAI Code, MAI Cyber and the media models are not stated by Microsoft on the pages checked. Among Phi models, Phi-4-mini-instruct and Phi-4-multimodal-instruct have 128K, Phi-4-reasoning has 32k, and Phi-4 and Phi-4-reasoning-vision-15B have 16K (16,384). See LLM Context Windows Explained.

Is any MAI or Phi model open source?

Phi models are open weight under the MIT licence on every Phi model card checked (Phi-4, Phi-4-mini-instruct, Phi-4-multimodal-instruct, Phi-4-reasoning, Phi-4-reasoning-vision-15B and Phi-Ground-Any). MAI models are not released as weights, and the microsoft.ai pages state no licence for them; they are accessed through Microsoft products, Foundry, the MAI Playground or named partners. See Open-Weight vs Closed LLMs Explained.

Model cards and system cards

See Model Cards and System Cards Explained.

Benchmarks

As of 15 November 2026. Microsoft's own scores for its own models; provider-reported and independent results are kept in separate columns.

See How to Read LLM Benchmarks.

Deprecations and retirements

As of 15 November 2026. Deprecation or retirement dates for MAI models: not stated by Microsoft on the pages checked (2026-11-15). The microsoft.ai models page now links newer versions (MAI Code 1.1 Flash, Image 2.6, Transcribe 2, Voice 2.1) in place of the versions announced at Build; whether the Build versions remain available is not stated on the pages checked.

Hosted Phi deployments on Foundry follow Foundry's own model retirement schedule; check the catalogue entry for the model you deploy. See the AI Model Retirement Guide.

Data handling and security

• Models sold by Azure in Foundry: Microsoft's Foundry data privacy page states that prompts, completions, embeddings and training data are not used to train generative AI foundation models without your permission or instruction, are not available to the providers of models sold by Azure, and that the models are stateless.
• Stored data: stateful features (Responses API, stored completions, files and vector stores, fine-tuning, batch) store data at rest in your Foundry resource within your Azure tenant and geography, encrypted with AES-256 by default, with customer-managed keys supported.
• Abuse monitoring: prompts and completions may be stored in a separate abuse monitoring store for human review. Managed customers can apply to modify abuse monitoring; when storage is off, a ContentLogging capability with value false appears in the resource's JSON view.
• Which policy applies: check whether a MAI model is listed as sold directly by Azure in the Foundry catalogue before relying on the policy above; the microsoft.ai pages do not state it.
• MAI Playground: personal, non-commercial use only; do not use it for work data.
• Secure coding: Microsoft states MAI Code 1.1 Flash was evaluated for secure coding on CyberBench, CyberSecEval and SecRepo.
• Phi self-hosting: data stays on infrastructure you control; the security of the serving stack is your responsibility.

Source: Microsoft Learn, Data, privacy and security for Microsoft Foundry models sold by Azure, accessed 2026-11-15. See the Azure OpenAI and Azure AI Foundry Security Guide and the LLMjacking Guide.

How to choose a Microsoft MAI or Phi model: factual criteria

• Weights: Phi models can be downloaded and self-hosted under MIT; MAI models are reached through Microsoft services or named partners.
• Size and hardware: Phi-4-mini-instruct is 3.8B and Phi-4-multimodal-instruct about 5.6B; Phi-4 and Phi-4-reasoning are 14B; MAI Thinking 1 has 35B active parameters.
• Context: 256k on MAI Thinking 1; 128K on Phi-4-mini and Phi-4-multimodal; 32k on Phi-4-reasoning; 16K on Phi-4.
• Modality: audio input on Phi-4-multimodal-instruct; image input on Phi-4-reasoning-vision-15B; speech models in MAI Transcribe and MAI Voice.
• Commercial use: the MAI Playground is personal and non-commercial; production use goes through Foundry, Microsoft products or partners.
• Access: MAI Cyber 1.1 Flash requires verification as a defender.
• Price data: published for MAI Transcribe 2 and MAI Voice 2.1 only; confirm others in Foundry before budgeting.

Practitioner checklist

• Confirm the exact model and version on microsoft.ai or the Hugging Face card before you build, since newer MAI versions replace Build versions on the models page.
• Check the price for your deployment route (Foundry, GitHub Copilot, partner platform) and record the date checked; note that the MAI Transcribe 2 price is promotional.
• Check the context window for the specific Phi variant, which ranges from 16K to 128K.
• Keep work data out of the MAI Playground, which is for personal, non-commercial use.
• Confirm whether a MAI model is sold directly by Azure before applying the Foundry data privacy terms to it.
• For self-hosted Phi, pin the repository revision, verify the files and scan the serving stack.
• Store Foundry and partner API keys in a secrets manager and rotate them; never embed them in client code.

Standards and references

• Microsoft AI: Models (accessed 2026-11-15)
• Microsoft AI: Introducing MAI Thinking 1 (accessed 2026-11-15)
• Microsoft AI: Microsoft Build 2026 MAI keynote transcript (accessed 2026-11-15)
• Microsoft AI: MAI Code 1.1 Flash (accessed 2026-11-15)
• Microsoft AI: MAI Cyber 1.1 Flash (accessed 2026-11-15)
• Microsoft AI: MAI Transcribe 2 (accessed 2026-11-15)
• Microsoft AI: MAI Voice 2.1 (accessed 2026-11-15)
• Microsoft AI: MAI Image 2.6 (accessed 2026-11-15)
• Microsoft on Hugging Face: phi-4 model card (accessed 2026-11-15)
• Microsoft on Hugging Face: Phi-4-reasoning-vision-15B model card (accessed 2026-11-15)
• Microsoft on Hugging Face: Phi-4-mini-instruct model card (accessed 2026-11-15)
• Microsoft Learn: Data, privacy and security for Foundry models sold by Azure (accessed 2026-11-15)

Related NHI Mgmt Group resources: Amazon Nova Models, Mistral AI Models, Small and Low-Cost LLMs Compared, Azure OpenAI and Azure AI Foundry Security Guide.