What it is
Phi-4-Reasoning-Vision-15B is the newest model in Microsoft's Phi family of small models, released on March 4, 2026. It combines the Phi-4-Reasoning language model with an image encoder, so it reads text and images and writes text. It has 15 billion parameters and a 16,384-token context.
One model handles both styles: it can reason step by step or answer directly, and you can force either mode. It is trained mainly on English and is not meant for multilingual use.
The weights are free under the MIT licence on Hugging Face and GitHub, and the model is also offered through Microsoft Foundry.
Our review
The awkward fact about Phi-4-Reasoning-Vision-15B is in its own model card. Against Qwen3-VL-8B-Instruct, a model about half its size, it is behind on eight of the ten benchmarks Microsoft lists, including OCRBench (76.0 against 89.2), MathVista (75.2 against 77.1) and ScreenSpot-v2 screen grounding (88.2 against 91.5). It edges ahead only on AI2D diagrams and ChartQA, by 2.1 points or less.
Where it does well is against older or general models. It is far ahead of Phi-4-multimodal and Gemma 3 12B on charts and screens; Gemma 3 12B scores 3.5 on ScreenSpot-v2. Microsoft compared it only with 2025-era models, not Gemma 4, and all of these scores are self-reported.
The 16,384-token context is the practical ceiling. The vision encoder can spend up to 3,600 tokens on a single high-resolution image, so a few screenshots and a reasoning trace exhaust the window. Microsoft recommends bf16 on vLLM, which means roughly 30GB of weights; it lists testing on A6000, A100, H100 and B200 GPUs, not consumer cards.
The MIT licence and the switchable reasoning mode are genuine advantages, and its published safety defect rates (1.4% text-to-text, 4.5% image-to-text) are more than most open releases offer. Adoption has not followed: the Hugging Face page logged about 3,100 downloads in the past month, which means little community quantisation or tooling to lean on.
Where it earns its place
A platform engineer at a company whose legal team approves only MIT or Apache licences
Add chart and diagram question-answering to an internal analytics tool
Serve it in bf16 with vLLM on one A100 or A6000 and use non-thinking mode for speed. It is a defensible choice on licence, but Qwen3-VL-8B scores higher on the same card.
A team prototyping a computer-use agent for legacy desktop software
Locate buttons and fields on screenshots
88.2 on ScreenSpot-v2 is serviceable for grounding, though the 16K context limits how many screens fit per step. Expect to cap history aggressively.
A product manager at an education start-up in the UK
Explain handwritten maths and diagram problems to students
Forced thinking mode raises MathVerse from 44.9 to 53.1, but costs latency and tokens. English-only training rules it out for multilingual markets.
Before you commit
- Setup
- Requires torch 2.7.1+, Transformers 4.57.1+ and, for serving, vLLM 0.15.2+; Microsoft says always use the chat template and a system prompt.
- Integrations
- Hugging Face and GitHub weights, Microsoft Foundry hosting, vLLM and Transformers; thinking or non-thinking output forced by appending a token to the generation template.
- Data and privacy
- MIT licence; self-hosting keeps images and prompts local, and Foundry hosting follows Azure's terms.
- Cost at scale
- Self-hosting needs a GPU with well over 30GB, such as a 48GB A6000 or an 80GB A100; Foundry pricing is not published on the model card.
- Limits
- 16,384-token context shared between images and text; English only; Microsoft lists it as unsuitable for medical, legal or financial advice.
What's good
- MIT licence with no usage restrictions
- Thinking and non-thinking modes in one model
- Published safety red-team defect rates
- Well ahead of Gemma 3 12B on charts and screens
What's not
- Behind Qwen3-VL-8B on 8 of 10 benchmarks in Microsoft's own table
- 16K-token context, with up to 3,600 tokens per image
- About 30GB in bf16; tested only on data-centre GPUs
- English only, by the model card's own statement
Best for
Teams with a data-centre GPU who need an MIT-licensed English vision model for charts, diagrams and screenshots and are prepared to compare it with Qwen3-VL first.
Alternatives
Phi-4-Reasoning-Vision-15B: common questions
- Is Phi-4-Reasoning-Vision-15B worth it?
- Phi-4-Reasoning-Vision-15B is an MIT-licensed vision model that loses to an 8B Qwen on Microsoft's own scorecard. We rate it 3.4 out of 5.
- How much does Phi-4-Reasoning-Vision-15B cost?
- Free weights under MIT; Microsoft Foundry hosting billed separately.
- What are the downsides of Phi-4-Reasoning-Vision-15B?
- Behind Qwen3-VL-8B on 8 of 10 benchmarks in Microsoft's own table; 16K-token context, with up to 3,600 tokens per image; about 30GB in bf16; tested only on data-centre GPUs; english only, by the model card's own statement.
- Who is Phi-4-Reasoning-Vision-15B best for?
- Teams with a data-centre GPU who need an MIT-licensed English vision model for charts, diagrams and screenshots and are prepared to compare it with Qwen3-VL first.
- How much does Phi-4-Reasoning-Vision-15B cost for a team or in production?
- Self-hosting needs a GPU with well over 30GB, such as a 48GB A6000 or an 80GB A100; Foundry pricing is not published on the model card.
- Is Phi-4-Reasoning-Vision-15B safe for company data?
- MIT licence; self-hosting keeps images and prompts local, and Foundry hosting follows Azure's terms.
- What are the best Phi-4-Reasoning-Vision-15B alternatives?
- The closest alternatives we have reviewed are Gemma 4, Mistral Small 4, Phi-4-multimodal and Nemotron 3.5 Super VL.
Answers checked Oct 22, 2026.