Aug 3, 2026

A local inference engine hand-built by antirez, the creator of Redis. Native code specialized for DeepSeek V4 Flash, targeting Metal first (Macs with 96 GB and up) plus CUDA and ROCm. Not a general GGUF runner: model loading through to the coding agent is designed and tested as one piece.

Why it matters: Bets against general-purpose runners, going narrow and deep on the best open weights only. A famous author entering local inference moves the engine race forward.

A measured writeup serving the 2.8T-parameter Kimi K3 at ~952 tok/s per MI355X node. The 1.5TB+ of weights no longer fit even a B200 node, and the authors argue AMD's larger HBM wins on performance per dollar.

Why it matters: As open models balloon, inference is starting to be decided by memory capacity rather than compute. Changes how buyers pick GPUs.

Gives your AI agent eyes on the whole internet: reading and search across X, Reddit, YouTube, GitHub, and Bilibili behind a single interface.

Why it matters: A sign that agent capability gaps now come from reachable information surface more than from the model. Includes Chinese-ecosystem sources from day one.