AI provider / Microsoft
Microsoft
A provider often discussed in connection with enterprise ecosystems, platform integration and compact model lines for practical deployment.
Microsoft connects cloud, productivity and enterprise infrastructure to applied AI at planetary scale — and increasingly ships its own models. The Phi family showed that carefully curated, "textbook-quality" data can pack large-model reasoning into phone-sized models.
Alongside Phi, the newer in-house MAI family marks a more direct first-party strategy, spanning text, voice, image and the MAI-Thinking reasoning line, delivered natively through Microsoft Foundry. Together they pair platform reach with compact, efficiency-minded models.
Model catalog
Every release in this provider’s model line.
Each row is a distinct model release — its date, its type and the traits it is known for. Open a source to dig into any of them.
| Model | Released | Type | Key traits | Source |
|---|---|---|---|---|
| Microsoft-Decision-1 | Oct 2026 | Decision |
| |
| MAI-Thinking-1 | Jun 2026 | Reasoning |
| |
| MAI in Foundry (Transcribe/Voice/Image) | Apr 2026 | Model family |
| |
| Phi-4-reasoning-vision | Mar 2026 | Reasoning |
| |
| MAI-1-preview & MAI-Voice-1 | Aug 2025 | Debut |
| |
| Phi-4-reasoning | Apr 2025 | Reasoning |
| |
| Phi-4 | Dec 2024 | Version update |
| |
| Phi-3 | Apr 2024 | Open release |
| |
| Phi-2 | Dec 2023 | Compact |
| |
| Phi-1 | Jun 2023 | Research |
|
Timeline
A practical timeline of model releases and company milestones.
This page tracks the moments that matter most: major model releases, version jumps and company moves that changed how this provider shows up in the AI market.
Microsoft finishes pre-training a 2.8-trillion-parameter model, MAI-Base-2.
Microsoft AI claims it beats every open-weight model tested while needing up to 12x less compute than a leading rival — the model is still in post-training, with no release date, pricing or public benchmarks yet.
Microsoft-Decision-1 trades conversation for calibrated choices.
Built on Qwen3.5-9B and tuned for routing, classification and verification, it answers 35x faster than GPT-6 Sol at a fraction of the cost — Xbox used it to triage over 10,000 pieces of player feedback.
MAI-Thinking-1 headlines seven new MAI models at Build.
A 35B MoE reasoner plus Code, Image, Voice and Transcribe updates turn MAI into a full first-party lineup overnight.
MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 land in Foundry.
The first output of the Superintelligence team makes MAI a commercial multi-modality family, not a Copilot-only experiment.
Phi-4-reasoning-vision decides for itself when to think.
The 15B vision-language model adds adaptive reasoning, while distilled Phi Silica variants ship inside Copilot+ PCs.
MAI-1-preview and MAI-Voice-1 reveal Microsoft’s in-house bet.
The first fully home-grown foundation models signal that Copilot’s future will not depend on OpenAI alone.
Phi-4-reasoning brings o-style thinking to small models.
After February’s mini and multimodal variants, the reasoning editions show deliberate chain-of-thought is not reserved for giants.
Phi-4 specializes the small-model line in complex math reasoning.
At 14B, synthetic-data training lifts it past much larger models on math competitions, sharpening Phi’s identity.
Phi-3 goes open weight and phone-sized.
Phi-3-mini runs on a smartphone while rivaling GPT-3.5-class quality, turning the research line into a deployable product family.
Phi-2 packs sub-13B-class reasoning into 2.7B parameters.
The follow-up makes the data-quality thesis hard to ignore, matching models up to 25x its size on reasoning benchmarks.
Phi-1 shows “textbook quality” data beats raw scale.
A 1.3B model trained on curated data outperforms far larger code models, founding Microsoft’s small-language-model program.