Phi (language model)

Phi
Phi
Developer	Microsoft
Release	June 21, 2023; 3 years ago
License	MIT
Website	azure.microsoft.com/en-us/products/phi/

Phi is a series of large language models developed by Microsoft that are open-weights and can run locally on one's device.

The initial version, Phi-1, released in June 2023. Phi-3 was released in April 2024. Phi-4-reasoning-vision-15B was released in March 2026.

The New York Times considered Phi-3-mini to be Microsoft's push into smaller large language models. Phi-3-mini has 3.8 billion parameters. Phi-3-mini has been compared to GPT-3.5 in terms of capability.

History

On June 21, 2023, Microsoft released Phi-1 with 1.3 billion parameters.^[2]

In April 2024, Microsoft released Phi-3-mini as an open-weights model with 3.8 billion parameters. Phi-3-mini is capable of running on a phone.^[3]

In March 2026, Phi-4-reasoning-vision-15B was released, with the capability of deciding when to reason.^[4]

Models

Phi-3-mini

Phi-3-mini was released in April 2024.^[5] It has 3.8 billion parameters and it was released into Azure. It is also capable of running on a phone.^[3]

The New York Times considered Phi-3-mini to be Microsoft's push into smaller language models.^[5] WIRED and Ars Technica considered Phi-3-mini to be of similar quality to GPT-3.5.^[6]

Phi-4-reasoning-vision-15B

Phi-4-reasoning-vision-15B was released in March 2026. It is capable of deciding when to provide a reasoning process before answering the user prompt, and when to skip reasoning. It has 15 billion parameters, and is a vision-language model.^[4]

References

↑ "LICENSE · microsoft/phi-4". Hugging Face. December 11, 2024. Retrieved April 12, 2026.
↑ Pandey, Mohit (June 21, 2023). "Microsoft Releases 1.3 Bn Parameter Language Model, Outperforms LLaMa". Analytics India Magazine. Archived from the original on November 22, 2024. Retrieved April 12, 2026.
1 2 Edwards, Benji (April 23, 2024). "Microsoft's Phi-3 shows the surprising power of small, locally run AI language models". Ars Technica. Retrieved January 20, 2026.
1 2 Jenic, Ivan (March 4, 2026). "Microsoft releases Phi-4-reasoning-vison-15B, an AI model that chooses when to think". Retrieved April 14, 2026.
1 2 Weise, Karen; Metz, Cade (April 23, 2024). "Microsoft Makes a New Push Into Smaller A.I. Systems". The New York Times. Retrieved January 20, 2026.
↑ Knight, Will (May 23, 2024). "Shrinking AI Programs Can Make Them More Powerful 🧠📲🤏". Fast Foward (Mailing list). WIRED. Retrieved January 20, 2026.{{cite mailing list}}: CS1 maint: url-status (link)

External links

Official website
Phi-4 on Hugging Face

[1] "LICENSE · microsoft/phi-4". Hugging Face. December 11, 2024. Retrieved April 12, 2026.

[2] Pandey, Mohit (June 21, 2023). "Microsoft Releases 1.3 Bn Parameter Language Model, Outperforms LLaMa". Analytics India Magazine. Archived from the original on November 22, 2024. Retrieved April 12, 2026.

[at-3-3] 1 2 Edwards, Benji (April 23, 2024). "Microsoft's Phi-3 shows the surprising power of small, locally run AI language models". Ars Technica. Retrieved January 20, 2026.

[nw-4-r-v-15b-4] 1 2 Jenic, Ivan (March 4, 2026). "Microsoft releases Phi-4-reasoning-vison-15B, an AI model that chooses when to think". Retrieved April 14, 2026.

[nyt-3-5] 1 2 Weise, Karen; Metz, Cade (April 23, 2024). "Microsoft Makes a New Push Into Smaller A.I. Systems". The New York Times. Retrieved January 20, 2026.

[Wired_Shrinking_AI_Programs-6] Knight, Will (May 23, 2024). "Shrinking AI Programs Can Make Them More Powerful 🧠📲🤏". Fast Foward (Mailing list). WIRED. Retrieved January 20, 2026.{{cite mailing list}}: CS1 maint: url-status (link)

[1]

[2]

[3]

[4]

[5]

[6]