New in 26.10
Alpamayo 1.5 Driving VLA
NVIDIA's Alpamayo 1.5 driving VLA on Chimera GPNPU: W4A16 vision tower, 8B reasoning model and action expert as ChiPy programs on four cores, checked against...
SDK 26.10Open the demo
Every network running on the Chimera GPNPU, newest and most complete first. 400+ models, demos, and building blocks, each with a runnable notebook or the full source graph behind it.
NVIDIA's Alpamayo 1.5 driving VLA on Chimera GPNPU: W4A16 vision tower, 8B reasoning model and action expert as ChiPy programs on four cores, checked against...
Runs the full 40-layer Qwen3.6 decoder end to end on the Chimera ISS via ChiPy — real token in, predicted token out — with the DeltaNet state and KV cache...
Ports Qwen3-VL-2B from PyTorch without ONNX: a W4A16 vision tower and language model as ChiPy chains of CCL kernels, describing a photo on the Chimera ISS.
Run Liquid AI's LFM2.5-350M on the Chimera ISS using imported QAD weights, 16-bit activations, and a PyTorch-native ChiPy decode pipeline.
Validates one Qwen3-8B prefill decoder layer against ONNX Runtime on a 4-core QC-P ISS and compiles the full 36-layer prefill graph with CGC.
Compiles the entire NMS-free YOLO26 detector into a single int8 program on the Chimera GPNPU, with CCL custom-op kernels for the PSA attention blocks and the...
Runs PaliGemma 3B mix-448 INT8 vision-language prefill as one fused CCL kernel on the Chimera GPNPU: SigLIP tail, projector, token bridge and a Gemma-1 decoder...
Sarashina2.2-3B quantized to W4A16 (SmoothQuant, int4 weight-only) and run as one Chimera kernel with custom attention and projection ops, gated against ONNX...
INT8 ModernBERT-base masked language model compiled for a single-core QC-N with 1 MB OCM through nine custom ops, validated on the ISS against ONNX Runtime.
Create a free DevStudio account to search the full catalog, open the demo notebooks, and see cycle-accurate benchmarks, source code, and intermediate representations for every model.