jul 12, 2026
2 links from the engineering internet.
github.com
aioss
llama.cpp b9970 adds a lightning-indexer op for deepseek v3.2/v4
llama.cpp build b9970 introduces a new ggml lightning-indexer operation that wires up deepseek v3.2 and v4's sparse-attention indexer, a prerequisite for running those models' native long-context path on the cpu and gpu backends.
github.com
aiinfra
litellm v1.92.0 ships day-0 gemini 3.5 flash and trims streaming overhead
litellm's stable v1.92.0 adds day-0 gemini 3.5 flash support, native mcp oauth for cursor with per-server rpm limits, and cuts about 30% of per-chunk streaming overhead on the anthropic and bedrock hot paths.