jul 3, 2026
4 links from the engineering internet.
github.com
aisystems
llama.cpp b9866 adds cuda topk-moe fusion for 288-expert models
the build fuses top-k mixture-of-experts routing into a single cuda kernel for 288-expert models like step-3.7, cutting fixed routing overhead for roughly a 2.4% decode speedup at shallow context depths.
github.com
oss
n8n 2.28.6 fixes duplicate zod instances that broke npm installs
the patch stops n8n's core package from pulling in duplicate zod copies, a clash that had been breaking npm installs, and realigns parameter input sizing in the editor.
github.com
databases
clickhouse ships v25.8.26.11 lts maintenance patch
clickhouse tags v25.8.26.11-lts, another maintenance patch on the 25.8 long-term-support line, shipping refreshed client and server binaries across architectures.
github.com
ai
litellm 1.90.2 backports gemini live fixes and cosign signing
the stable patch backports fixes for gemini live setup and hung-handshake recovery, and adds docker image signature verification through cosign.