Chat with a model that never leaves your machine.

llama-wgsl downloads SmolLM2 once, streams the weights straight into GPU memory, and runs every layer with compute shaders written by hand in WGSL. Pick a model and press Load.

Model

checking WebGPU…

Checkpoint

Speed

last reply

—tokens / s

prompt: –
context: –
weights: –
KV cache: –
dispatches: –
adapter: –

Sampling

Temperature 0.7 Top-p 0.9 Repetition penalty 1.1 Max new tokens 256

What runs per token

embed gather one row
× layers
1. rmsnorm
2. qkv one fused matvec
3. rope + write K/V cache
4. attention GQA, a workgroup per head
5. o_proj + residual add
6. rmsnorm
7. gate·up fused with SiLU
8. down + residual add
lm_head tied embedding matvec
sample top-k / top-p on the CPU

All dispatches for a token go in one compute pass. Weights stay bf16 on the GPU and are widened to f32 in registers.