2026-10-11 18:00 UTC

moe-models

band: coolmomentum: stable score: 0.271
temperature history

Episodes (5)

Independent benchmarks will determine whether HotPin's llama.cpp patches can run 30B–120B MoE models losslessly on roughly 24GB of consumer RAM at practical speeds without excessive storage wear.
expirednovelscott: none
Independent evaluation will determine whether the reported 92GB one-bit quantization of Tencent's Hy3 295B preserves coding quality while outperforming its cloud API on a four-RTX-5090 system.
expiredconvergesscott: medium
Independent deployments will determine whether Lumabri and Colibri can serve mixture-of-experts models across peer-to-peer commodity machines with practically useful throughput and reliability.
expiredknownscott: medium
Independent reproduction will determine whether the reported train-inference mismatch in open-weight MoE reinforcement-learning stacks is a widespread failure mode that materially undermines reproducibility and deployed-model quality.
expiredconvergesscott: medium
LocalLLaMA builder Ok-Breadfruit-3523 claims a ~$800 rig of five ex-mining BC-250 boards exposes ~71GB VRAM and serves Qwen3-Coder-Next Q4 at 40 tok/s (30k context, ~30 at 100k) over 1GbE with headless worker boards — replication or wider adoption of ex-mining BC-250 rigs would establish salvaged mining hardware as a cheap ~70GB route to local large-MoE coding inference despite its power inefficiency.
watchingconvergesscott: medium

Trajectory notes