Phi-4-reasoning-vision-15B

What it is

Phi-4-reasoning-vision-15B is a vision-language model with 15 billion parameters, released with open weights by Microsoft on 4 March 2026^[1][2]. It takes text and images as input and returns text, and it works within a context length of 16,384 tokens^[1]. Open weights means the trained parameters are published, so the model can be downloaded, run and adapted locally instead of only being reached through a hosted service; the release uses the MIT licence^[1].

How it is put together

The model combines two parts. One is the Phi-4-reasoning language backbone, which handles language and reasoning; the other is a SigLIP-2 vision encoder, which turns an image into representations the language side can read^[1]. They are joined in a mid-fusion design in which visual tokens are projected into the language model, so the picture and the text are handled in one stream rather than by separate models stitched together at the end^[1]. Images are processed at dynamic resolution, producing up to 3,600 visual tokens, so a dense chart or a screenshot can be represented in enough detail to be read rather than collapsed into a coarse summary^[1].

Inside each image the model applies bidirectional attention, letting one part of the image attend to parts on either side of it instead of only to what came before^[1]. The stated purpose was to improve spatial reasoning, and the design avoided the overfitting that appeared when bidirectional attention was applied more broadly across the model^[1].

Deciding when to reason

Reasoning models often spend extra computation on intermediate steps, which helps on hard problems but is wasted on simple ones. This model makes that choice itself: it uses `<think>` blocks, the step-by-step working, for math and science questions, and `<nothink>` for direct answers on perception tasks such as captioning an image or grounding an object in it — that is, saying where in the picture something is^[1][2]. Users can also force either mode rather than leaving the decision to the model^[1][2].

How it was trained

Training was supervised fine-tuning, the process of showing the model pairs of inputs and desired outputs and adjusting it toward them, on a mix of reasoning and non-reasoning data^[1]. The data was mainly filtered open-source vision-language datasets, supplemented by domain data from internal Microsoft teams and by targeted acquisitions^[1]. The run took 4 days on 240 NVIDIA B200 GPUs^[1].

Safety alignment used further supervised fine-tuning on open-source and in-house synthetic data containing helpfulness and harmlessness examples, teaching the model to refuse hate speech, violence, self-harm and sexually explicit content^[1].

Results

In its default mode the model scores 84.8 on AI2D, 83.3 on ChartQA, 75.2 on MathVista-mini, 44.9 on MathVerse-mini and 88.2 on ScreenSpot v2^[1]. These cover diagram and chart comprehension, math problems posed over images, and locating elements on a screen; the spread is a reminder that reading a chart and solving a diagram-based math problem are different difficulties for the same model.

What it is for

Microsoft names two target uses: scientific and math reasoning over images, and computer-use agent tasks such as locating GUI elements — the buttons, fields and icons an agent must find before it can click them^[1]. Both depend on the same ability, connecting a question in words to a specific region of a picture, and the choice to skip step-by-step working on perception tasks keeps those calls cheap. With open weights and an MIT licence, the model is positioned for teams that want a mid-sized vision-language model they can inspect, fine-tune and run themselves.