Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
ci: kimi_k3.exe with the DLL loader too, it links the same backend
JustVugg0e3272ae7d05
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.ymlindex 3579a739..906804e5 100644--- a/.github/workflows/ci.yml+++ b/.github/workflows/ci.yml@@ -827,6 +827,12 @@ jobs:make qwen36.exe CUDA_DLL=1test -f qwen36.exe || { echo "qwen36 CUDA_DLL=1 reported success but produced no exe" >&2; exit 1; }echo "qwen36.exe built against the DLL loader"+ # kimi_k3 links the same loader: three engines share coli_cuda.dll.+ # deepseek_v4 has its own DLL (coli_cuda_dsv4.dll, Makefile.deepseek-v4)+ # and joins in a later step; inkling's CUDA object has no Windows loader.+ make kimi_k3.exe CUDA_DLL=1+ test -f kimi_k3.exe || { echo "kimi_k3 CUDA_DLL=1 reported success but produced no exe" >&2; exit 1; }+ echo "kimi_k3.exe built against the DLL loader"- name: Collect the runtime the DLL links (cudart64_12.dll)shell: bashrun: |@@ -846,6 +852,7 @@ jobs:c/cudart64_*.dllc/colibri.exec/qwen36.exe+ c/kimi_k3.exeweb:name: Web UI