Transformers.js v4 hits npm: WebGPU comes to Node.js with 4x faster models

Xenova dropped Transformers.js v4 on npm today, and this isn’t a minor polish job. It’s a full-blown architectural overhaul that finally makes WebGPU a first-class citizen across every JavaScript environment — not just the browser. The team worked directly with the ONNX Runtime crew to build a new WebGPU runtime in C++, which now runs in Node, Bun, and Deno. That means server-side JavaScript developers get hardware-accelerated inference without leaving their runtime, a shift that quietly erases a major friction point for local AI deployment.

The performance numbers aren’t theoretical either. By ripping out generic ONNX ops and replacing them with specialized contrib operators like com.microsoft.MultiHeadAttention, the team clocked a 4x speedup for BERT embedding models. For larger language models, they’re leaning on operators like GroupQueryAttention and MatMulNBits, re-implementing model logic operation by operation to squeeze out every bit of speed in resource-constrained environments. The result is state-of-the-art models running 100% locally, but now at speeds that feel less like a tech demo and more like a tool you’d actually use.

Under the hood, the repo’s been split into a pnpm monorepo, the monolithic 8,000-line models.js file is dead and buried, and the build system jumped from Webpack to esbuild — cutting build times from 2 seconds to 200 milliseconds while shipping bundles that are 53% smaller. New architectures like Mamba, Multi-head Latent Attention, and Mixture of Experts models are all WebGPU-compatible, with names like Olmo3, FalconH1, and HunYuanDenseV1 now available. The ModelRegistry API also arrives for production use cases, letting developers manage model loading and caching with more control. The message is clear: browser-based AI just got a backend runtime, and it’s actually fast.