Model Playground
Run machine-learning models in your browser, on your own GPU — speech recognition, object detection, translation, text generation and 40 more tasks across audio, vision, language, graphs, tabular data and reinforcement learning. The weights come to you; nothing you give a model is sent to a server (privacy).
Audio
- Text to Speech — Turn text into speech in your browser with Kokoro, MMS and SpeechT5 — synthesised on your own GPU or CPU, with nothing sent to a server.
- Automatic Speech Recognition — Transcribe speech to text in your browser with Whisper and Moonshine — live from the microphone or from a file, privately, on your own device.
- Audio to Audio — Remove background noise from speech in your browser with DeepFilterNet3 — full-band 48 kHz enhancement that never uploads your recording.
- Audio Classification — Tag sounds in your browser with AST, wav2vec2 and CLAP — fixed labels or your own text prompts, scored on your device with nothing uploaded.
- Voice Activity Detection — Find where speech starts and stops in a recording with Silero VAD, running in your browser on the CPU at around 100x real time.
Computer Vision
- Depth Estimation — Estimate depth from a single photo in your browser with Depth Anything — a relative depth map computed on your own GPU, never uploaded.
- Image Classification — Classify photos in your browser with ViT, ResNet and MobileNet — from a file or your webcam, with ranked labels computed on your device.
- Object Detection — Detect and box objects in photos or a live webcam feed in your browser with D-FINE, RT-DETR, YOLOS and DETR running on your own GPU.
- Image Segmentation — Segment images into labelled regions in your browser with SegFormer and DETR panoptic — masks computed on your device, with an overlay.
- Image to Image — Upscale images with Swin2SR super-resolution in your browser, tiled so large pictures fit, and compared against a bicubic baseline.
- Video Classification — Classify video clips in your browser by sampling frames through an image model — a frame-level baseline running on your own device.
- Zero Shot Image Classification — Classify images against labels you type with CLIP and SigLIP in your browser — no training, and the label embeddings are cached.
- Mask Generation — Click on an image to segment any object with Segment Anything (SAM) in your browser — encode once, then decode a mask per click.
- Zero Shot Object Detection — Detect objects by name with OWL-ViT and OWLv2 in your browser — type what to look for and get boxes without training a detector.
- Image to 3D — Turn a single photo into an interactive 3D point cloud in your browser — Depth Anything depth, unprojected and rendered with WebGPU.
- Image Feature Extraction — Extract image embeddings with CLIP, DINOv2 and DINOv3 in your browser and compare pictures by cosine similarity, on your own device.
- Keypoint Detection — Estimate human pose in your browser — detect each person, then place body keypoints and a skeleton, from a photo or your webcam.
- Background Removal — Remove the background from a photo in your browser with MODNet or RMBG — a soft alpha matte and a transparent PNG, never uploaded.
Multimodal
- Image Text to Text — Ask questions about an image with SmolVLM, a vision-language model running in your browser on WebGPU, with the answer streamed as it is written.
- Visual Question Answering — Visual question answering in your browser — ask a short question about a picture and get a one-word answer from a small VLM on WebGPU.
- Video Text to Text — Ask a vision-language model about a video in your browser — SmolVLM2 reads frames sampled from the clip on WebGPU and answers in text.
Natural Language Processing
- Text Classification — Classify text in your browser — sentiment and financial-tone models side by side, with the full score list, running on your own device.
- Token Classification — Find names, places and organisations with named-entity recognition in your browser, and redact them without uploading the document.
- Question Answering — Extractive question answering in your browser — paste a passage, ask a question, and see the exact answer span highlighted in the text.
- Zero Shot Classification — Classify text against labels you type with an NLI model in your browser — no training, an editable hypothesis template, one pass per label.
- Translation — Translate text with Marian models in your browser — one language pair per model, private machine translation on your own device.
- Summarization — Summarize articles in your browser with T5 and DistilBART, measured against a lead-3 baseline, with a sentence-level faithfulness check.
- Feature Extraction — Turn sentences into embeddings in your browser with MiniLM, BGE and Nomic models — inspect the vectors and try Matryoshka truncation.
- Text Generation — Generate text with small language models such as SmolLM2 and GPT-2 in your browser — compare greedy decoding with sampling, streamed live.
- Fill Mask — Predict a masked word with BERT, RoBERTa and ModernBERT in your browser — and probe what each model's training corpus taught it.
- Sentence Similarity — Compare sentences by meaning in your browser — embedding models score paraphrases against unrelated text with cosine similarity.
- Text Ranking — Rank documents for a query in your browser with BM25, dense embeddings, hybrid fusion and a cross-encoder reranker, side by side.
Other
- Graph Machine Learning — Train a graph neural network (GCN, GraphSAGE, GIN, GAT) on the Cora citation graph in your browser with WebGPU, and watch oversmoothing.
- Link Prediction — Predict missing citations in the Cora graph with a GNN trained in your browser on WebGPU, scored by AUC on held-out edges.
- Graph Classification — Classify protein graphs from the PROTEINS dataset with a graph neural network trained in your browser, beside a majority baseline.
Reinforcement Learning
- Reinforcement Learning — Watch Q-learning, REINFORCE and Actor-Critic learn FrozenLake and CartPole in your browser, checked against exact value iteration.
- Robotics — Robot perception in your browser — find objects by name with OWLv2 and order them by depth — plus a behaviour-cloning experiment.
Tabular
- Tabular Classification — Fit classifiers to your own CSV in your browser — logistic regression, random forest, gradient boosting and an MLP, with nothing uploaded.
- Tabular Regression — Fit regression models to your own CSV in your browser — ridge, trees, boosting and quantile bands, with residual plots and nothing uploaded.
- Time Series Forecasting — Forecast a time series in your browser with naive, seasonal and drift baselines, scored by a rolling-origin backtest and MASE.
Theory
- Linear Model Training — Train a softmax classifier on MNIST in your browser with hand-written WebGPU kernels, and watch its weights turn into digit templates.
- Discrete Maths — Step a breadth-first search and count walks with powers of the adjacency matrix — why a k-layer graph network reads only k hops.
- Tensor Arithmetic — Run matrix and tensor operations on your GPU with raw WebGPU compute shaders in a Web Worker — edit the operands and compute.
- GPU Playground — Pick a model and run it on custom WebGPU kernels in your browser — the raw-WGSL inference workspace behind the task pages.
Notes
- Fix “Missing required scale” in ONNX Runtime Web — Why quantized decoders fail on ONNX Runtime Web's WASM backend with qdq_actions.cc “Missing required scale”, and the per-graph dtype fix.
- Transformers.js NER and QA return no offsets — Transformers.js token-classification and question-answering leave start and end undefined. How to recover character offsets safely.
- WebGPU shader-f16: check it before loading q4f16 — “Program Gather requires f16 but the device does not support it”: why having a GPU is the wrong check for fp16 and q4f16 models.
- Send Transformers.js tensors through postMessage — Why a Transformers.js Tensor “could not be cloned” and a RawImage loses its methods in a Web Worker, and how to serialise both.
- When int8 quantization is wrong, not just worse — Measured cases where a q8 ONNX model gives a confident wrong answer in the browser, and how a known-answer test catches it.