Hikari_localLLM

Unofficial DeepSeek Harness plugin: mobile-friendly UI, Web Push notifications, and Linux Tailscale setup guide

No project description available.

Quant-native model surgery: edit packed quant weights without BF16 requant (Ternary Q2 worked example)

Serve DeepSeek-V4-Flash (NVFP4) on 2x RTX PRO 6000 (SM120) with working MTP speculative decoding (+38% single-stream). Root-cause fix: MTP draft experts are MXFP4, routed to Mxfp4MoEMethod.

Hikari's curated English measured notes on local LLM systems (serving, specdec, training, steering).

Truthful scheduler-step and streaming visualization for local LLM inference