📡 AINEWS RADAR 2026-10-11 16:37 UTC

moe

band: coolmomentum: stable score: 0.001

temperature history

Episodes (2)

Independent benchmarks will determine whether dsv4-streaming can run the 284B DeepSeek V4 Flash checkpoint on 64GB Apple Silicon with near-lossless quality and practically useful expert-streaming throughput.

expiredknownscott: medium

The fork author claims selectively placing frequently used MoE experts in VRAM raises llama.cpp generation throughput from 20 to 30 tokens per second on partially offloaded coding workloads, potentially improving local inference on memory-constrained GPUs.

resolvedknownscott: medium

Trajectory notes

2026-08-29T21:30:55Z: llama-cpp-hot-expert-offload closed (superseded) — The radar already tracks this exact development in `radar:llama-cpp-hot-expert-gpu-cache`. It bears directly on Scott’s hardware-aware placement policy and memory-constrained `gamepc` inference stack, but the claimed gain remai