Phi-4 Reasoning
Big Reasoning Power, Small Models
320 followers
Big Reasoning Power, Small Models
320 followers
Small open-weight models (3.8B/14B) delivering powerful reasoning for math/science/code, rivaling larger LLMs. Available on Azure AI Foundry & HF.
This is the 2nd launch from Phi-4 Reasoning. View more
Phi-4-reasoning-vision
Launching today
Phi-4-reasoning-vision-15B is a compact open-weight multimodal model built on a mid-fusion architecture. Balancing fast direct perception with deep chain-of-thought, building capable computer-use agents and solving complex math is now highly efficient.






Free
Launch Team

Flowtica Scribe
Hi everyone!
Phi-4-Reasoning-Vision-15B is Microsoft"s new 15B open-weight model that makes multimodal reasoning feel much more efficient.
It was trained on 200B multimodal tokens, handles high-res screens well, and stays direct on simpler tasks while switching into deeper reasoning when needed.
Looks especially strong for math, science, and computer-use agents. Weights on HF.
@zaczuo 15B with mid-fusion is a sweet spot — large enough for real reasoning but still runnable on a single 24GB card. The "direct perception vs deep chain-of-thought" switching is interesting. Does it decide that automatically based on task complexity, or is there a way to force one mode over the other?
@zaczuo @alan_silverstreams Mid-fusion in a 15B open-weight model is what makes Phi-4-Reasoning-Vision-15B interesting for GUI agents. Fast screen reads stay cheap, then you spend the extra reasoning only on the steps that need it.
The GUI agent angle is what makes this really compelling. A 15B model that can handle high-res screens well enough for computer-use tasks is a big deal for anyone building browser automation or testing tools. The adaptive reasoning depth -- going direct on simple perception but switching to chain-of-thought for harder problems -- seems like the right tradeoff for latency-sensitive agent loops. Have you seen benchmarks on how it compares to larger models specifically on GUI grounding tasks?