hjLabs AI Playground
Starting the on-device runtime…
transformers.js runs ONNX graphs through ONNX Runtime Web and covers 25 task types — vision, audio, embeddings, classification, and text generation. WebLLM compiles models with Apache TVM and MLC and specialises in chat: faster token generation for larger LLMs, with a smaller model selection and no vision or audio pipelines. This playground uses transformers.js for the twenty non-chat surfaces and evaluates WebLLM specifically for the chat page, where its throughput advantage on multi-billion parameter models is real.
transformers.js vs webllmonnx runtime web vs mlcbest browser llm librarywebgpu llm frameworkmlc web-llmhuggingface transformers jsin browser inference library
Yes, and it is often the right call — WebLLM for chat, transformers.js for everything else. They keep separate caches and separate WebGPU contexts, so budget the disk and VRAM for both.
transformers.js by a wide margin across tasks; WebLLM within chat models specifically, where its prebuilt library includes larger quantized Llama, Qwen and Phi builds.