NeuralStackMicrosoft started deploying Phi-4 a 3.8 billion parameter model to Windows 12 Copilot+ PCs on July 19 2026. The model runs at 42 tokens per second on Snapdragon X Elite processors. Users gain offline code completion and local image editing features.
Phi-4 was trained on 1.2 trillion tokens of synthetic and curated data. Microsoft claims 78 percent accuracy on HumanEval coding benchmarks. The rollout covers 18 million eligible devices by August 15.
Windows 12 launched in March 2026 with mandatory NPU requirements of 45 TOPS. Earlier Copilot+ features relied on cloud processing for complex tasks. Phi-4 reduces latency by 300 milliseconds on average.
Qualcomm and AMD collaborated on optimized inference kernels for the model. Microsoft open sourced the Phi-4 weights under a research license on July 20.
Why this matters
The move demonstrates viable on-device AI for mainstream laptops and reduces cloud dependency. Developers gain new local APIs for privacy-sensitive applications. Apple and Google face pressure to match performance on their hardware.
Microsoft expects 60 percent of Windows AI workloads to run locally by 2027. The strategy also lowers Azure inference costs significantly.
Future updates will add voice synthesis and real-time translation to the on-device suite.