Ready to load model (first run streams cleanly into browser OPFS)
Detecting WebGPU...

Ignis · Browser-Native WebGPU LLM & Full-Omni Engine

Built from first principles without ONNX, TVM, or llama.cpp: 256B-aligned .ignis shards, OPFS SyncAccessHandle row streaming, W3C baseline WGSL mega-kernels, Rust/WASM JSON Schema FSM, and Dual-Mode Speculative Decoding.

Spark-X2.6-4B (RX 6800 XT)
85.4 tok/s · 134 tok/s N-gram
LFM3-Omni-2B Full Omni
1.05 GiB · 15/15 PyTorch Parity
Gemma-4-E4B (262K + PLE)
2.69 GiB PLE Streamed via OPFS
WASM JSON Schema FSM
0.285 ms/tok · 6/6 Strict Match

Pure In-Browser WebGPU · Multi-Arch & Full-Omni

Streams weights and multi-gigabyte embeddings through OPFS directly into WebGPU inside a Dedicated Worker. Load LFM3-Omni-2B (Vision + Audio + 24 kHz Voice), Spark-X2.6-4B (85–134 tok/s), or Gemma-4-E4B-it, or click any Showcase button on the left.