#moe — HitReader

machine learning Jul 16, 2026 7 min read

Inkling and the new era of open-weights customization

Inkling is a multimodal open-weights MoE model (975B total, 41B active) built for customization: up to 1M-token context, controllable thinking effort, and fine-tuning via Tinker (LoRA). The release also demonstrates a self-finetuning loop that defines an objective, trains, evaluates, and swaps weights.

by ahsan

#fine-tuning #inkling #llm systems #moe #tinker

ai infrastructure Jul 10, 2026 7 min read

Run GLM-5.2 on a Slow PC with colibrì (Disk-Streamed MoE)

colibrì runs GLM-5.2 (744B MoE) on consumer hardware by keeping ~9.9GB of dense int4 weights in RAM and streaming routed experts from a ~370GB int4 container on disk. It uses an LRU expert cache, MLA-style compressed KV caching, and native MTP speculative decoding to improve interaction speed once caches are warm.

by ahsan

#cpu inference #glm-5.2 #moe #quantization #systems