Observed arrival · 2026-11-10

Barx: an 80B model in 5.5 GB

barxcompute.com Visit website

Editorial interest 78/100 Selection signal · not a rating of the site

Barx is a local inference engine that streams mixture-of-experts models from storage instead of loading the whole model into memory.

Landing page captured for the 2026-11-10 issue.

For: Developers running local language models on Apple silicon
Worth noticing: Its comparison uses a cold-start 45 GB Qwen3-Next-80B-A3B file and stock llama.cpp on the same CPU-only machine.

Field notes

Barx describes treating memory as a cache for model experts rather than a home for the entire model. Its stated expert budget is 2–4 GB, sized to the machine, and includes the engine, its workers, and model pages in the file cache. The page also says its kernels reproduce reference results bit for bit and that existing Ollama-compatible apps can connect unchanged.

Observed signals

Read the marks

Editorial observations of this landing page, not a rating.

○OpenPublic substance visible

✦PrettyNotable craft visible

●ProPolished or operationally mature

◎NicheUnusually specific use

One card from the complete issue

Detergent Lessons in Sheep Suits

328,969 arrived 1,000 judged 984 catalogued Enter the complete issue