Posted in

Microsoft Research announces Phi-4-reasoning-vision-15B m…

by ailona•

Phi-4-reasoning-vision-15B is a 15B open-weight multimodal reasoning model prioritizing efficiency, mid-fusion design, and curated multimodal training (~200B tokens). It uses SigLIP-2 Naflex, excels at math, science, and UI grounding, and advances the accuracy–compute Pareto frontier.

Phi-4-reasoning-vision-15B is a newly released 15B parameter open-weight multimodal reasoning model. It balances reasoning capability, efficiency, and moderate training data requirements.

Main feature/change and impact

Phi-4-reasoning-vision-15B uses a mid-fusion architecture with the SigLIP-2 Naflex vision encoder. The design choice prioritizes cross-modal reasoning while keeping compute and memory low. It achieves strong math and science reasoning and GUI grounding compared to similar-size models. The model pushes the pareto-frontier between accuracy and inference cost for practical deployment scenarios.

Practical implications

Developers gain a broadly capable VLM that runs on modest hardware with lower token budgets. The model was trained on 200B multimodal tokens, reducing training cost relative to trillion-token alternatives. Dynamic-resolution vision encoding improves performance on high-resolution and GUI-heavy inputs. Release artifacts are available on Microsoft Foundry, Hugging Face, and GitHub for direct use and fine-tuning.
“Phi-4-reasoning-vision-15B is a 15 billion parameter open‑weight multimodal reasoning model, available through Microsoft Foundry, HuggingFace and GitHub.”
Training lessons highlight architecture, data curation, and dataset composition as key levers. The team found that dynamic-resolution encoders outperform multi-crop and S2 variants on dense, high-resolution benchmarks. Data quality focused on filtered open-source datasets, domain-specific internal sets, and targeted acquisitions. Mixing reasoning-focused and perception-focused tasks yielded better generalization for multimodal reasoning. Closing: Organizations can adopt Phi-4-reasoning-vision-15B for image captioning, document reading, GUI understanding, and scientific problem solving. Next steps are integration tests, domain-specific fine-tuning, and benchmarking against deployment workloads.

Key points from the article:

  • Mid-fusion architecture balances performance with moderate compute requirements.
  • SigLIP-2 Naflex dynamic resolution boosts high-resolution image understanding.
  • Trained on curated multimodal data (~200B tokens) for data efficiency.
  • Strong performance on math and science multimodal reasoning tasks.
  • Offers competitive accuracy versus models requiring much more compute.

    • Related Coverage:

    From the Source