Microsoft Phi logo
Microsoft Phi logo

Microsoft Phi

Microsoft

MODEL DIRECTORY

MODEL DIRECTORY

Microsoft Phi

Microsoft Phi

Phi-4-reasoning-vision-15B is a compact open-weight multimodal specialist for math, science, charts, and UI grounding. Microsoft’s published results are strong for its size, but they remain provider-run and do not make Phi a broad frontier generalist.

CURRENT MODEL SNAPSHOT

Provider: Microsoft Research
Current anchor: Phi-4-reasoning-vision-15B
Lifecycle: Current
Weights: Open-weight
Reviewed: 21 July 2026

Abstract blue and purple gradient

Phi’s advantage is specialization per parameter

Microsoft released Phi-4-reasoning-vision-15B on 4 March 2026. The 15-billion-parameter open-weight model combines text and image reasoning, with particular emphasis on mathematics, science, charts, diagrams, and user-interface grounding. It can switch between reasoning and non-reasoning behavior rather than forcing long traces for every input.

Phi continues Microsoft’s small-language-model strategy: concentrate training and data quality so that a compact model can perform tasks that normally require a much larger checkpoint. The practical payoff is a smaller memory footprint, lower serving cost, and more realistic private or edge-adjacent deployment.

Compact does not mean universally capable. The model’s strongest evidence is domain-shaped, and its UI understanding should not be confused with permission to execute arbitrary GUI actions without controls.

The published results are strong for 15B, but provider-run

Microsoft reports 84.8 on AI2D, 83.3 on ChartQA, 64.4 on HallusionBench, and 44.9 on MathVerse in its reruns. The team published technical detail and evaluation logs, improving inspectability, but the measurements are still produced by the model provider.

Those tasks align with the product thesis: structured visual reasoning, charts, diagrams, and mathematical content. They do not establish frontier performance on open-ended research, long-running agents, multilingual operations, or broad business judgment.

  • The model should be compared with other compact multimodal systems, not only 100B-plus frontier models.

  • UI grounding must be tested for coordinate accuracy, stale-state recovery, and destructive-action prevention.

  • Reasoning and non-reasoning modes should be benchmarked separately for latency and cost.

The enterprise case: a focused model for visual reasoning

Phi is compelling for chart and diagram analysis, UI understanding, scientific documents, and private multimodal workflows where a 15B model is operationally attractive. It is not the right sole model for every enterprise agent; routing specialized visual tasks to Phi can be more credible than asking it to replace a broad frontier model.

How we would evaluate it

Build a test set from real charts, scanned tables, interfaces, and diagrams, including adversarial labels and changed UI state. Measure task accuracy, hallucinated elements, coordinate errors, recovery, latency, memory, and reviewer effort. Require approval before any consequential action.

Evidence used

Beam AI support status

Under evaluation. This page does not confirm a Beam integration, supported runtime, or permission model for GUI actions.

Use case 1

Chart and diagram analyst

Test real business charts, scientific figures, and diagrams with misleading legends and missing labels. Score numerical accuracy, unsupported claims, citations to visual regions, latency, and review time.

Use case 2

UI grounding assistant

Evaluate understanding of application screenshots and forms under changed layouts, disabled controls, and stale state. Measure element identification, coordinate accuracy, recovery, and blocked destructive actions.

Use case 3

Private scientific document review

Run the 15B model on controlled technical documents with equations and images. Track extraction, reasoning accuracy, omissions, hallucinations, memory use, and reviewer corrections.

Use case 4

Compact-model router

Route visual reasoning tasks to Phi and broader research to another model. Measure routing accuracy, end-to-end completion, blended latency and cost, and whether specialization reduces review debt.

Related LLMs

Curated alternatives to compare before selecting a model family.

ChatGPT

ChatGPT is OpenAI’s conversational product; enterprise automations use the underlying GPT models and APIs. This guide separates the product from the model family and explains how to evaluate the current GPT-5.6 lineup.

ChatGPT

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

DeepSeek

DeepSeek’s current V4 family provides Pro and Flash API models with thinking and non-thinking modes, a documented one-million-token context window, and compatible API interfaces.

DeepSeek

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Perplexity

Perplexity’s Sonar API family combines language models with web search and citations. The core enterprise decision is whether provider-managed retrieval and source handling fit the workflow.

Perplexity

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Gemini

Google’s Gemini family covers production and preview models for multimodal, reasoning, audio, media, and tool-oriented workloads. Model stage and endpoint stability are central to an enterprise selection.

Gemini

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Claude

Anthropic’s Claude family spans frontier reasoning, coding, vision, and faster production models. The practical choice is a specific pinned model ID on the Claude API or a supported cloud platform.

Claude

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Llama

Meta’s Llama 4 family provides open-weight, natively multimodal models. Scout emphasizes very long context and efficient deployment; Maverick targets stronger general multimodal capability.

Llama

Combine sections from a range of categories to easily assemble pages that meet the needs of your growing business.

Start Today

Build AI agents with the right model

See how Beam can orchestrate governed AI workflows across the model family that fits your requirements.

Start Today

Build AI agents with the right model

See how Beam can orchestrate governed AI workflows across the model family that fits your requirements.

Start Today

Build AI agents with the right model

See how Beam can orchestrate governed AI workflows across the model family that fits your requirements.

FAQs

Frequently Asked Questions

Model selection, deployment, governance, and Beam support questions answered.

What is Phi-4-reasoning-vision-15B?

It is Microsoft’s compact 15B open-weight model for multimodal reasoning, especially math, science, charts, diagrams, and UI grounding, with reasoning and non-reasoning behavior.

How strong are Phi’s visual benchmarks?

Microsoft reports 84.8 on AI2D, 83.3 on ChartQA, 64.4 on HallusionBench, and 44.9 on MathVerse. These are provider-run results with published technical context, not independent production evidence.

Is Microsoft Phi open-weight?

Yes. The model weights are published for download. Verify the exact license, repository, model card, notices, runtime dependencies, and artifact checksum before deployment.

Can Phi control enterprise software safely?

Visual grounding can support UI agents, but it does not make autonomous action safe. Test coordinates, stale state, permission boundaries, confirmation, destructive-action blocking, logging, and recovery.

Does Beam support Phi-4 reasoning vision in production?

Beam support is currently Under evaluation. No approved runtime, GUI-control pattern, or Beam benchmark is attached to this record.