Skip to content

Anubis OSS v3.8.0

Choose a tag to compare

View all tags
@uncSoft uncSoft released this 04 Jun 14:56
· 48 commits to main since this release
v3.8.0
0dfe951

Anubis OSS v3.8.0 — first-class oMLX support

This release makes oMLX a first-class backend in Anubis — accurate server-reported metrics, in-app model management, a built-in HuggingFace model browser, and a thinking toggle — plus a dark UI refresh and a build-from-source fix.

oMLX, front and center

  • First-class oMLX backend — a stock backend on port 8000 (separate from mlx-lm on 8080), with its own process detection, power/memory attribution, and owned_by: omlx identity verification.
  • Accurate, server-reported metrics — Anubis reads oMLX's own decode-loop numbers (generation tok/s, prefill tok/s, time-to-first-token, prompt-eval & generation duration, total time, model-load time, cached tokens) instead of estimating from wall-clock. No more occasional wildly-inflated tok/s or 0.00 J/tok.
  • Anubis × oMLX toggle in Session Details — flip between Anubis's hardware-measured view and oMLX's exact server-reported figures (shown verbatim, and saved with the run).
  • Model management built in — load and unload oMLX models on demand from the Vault (with real sizes and loaded state) and via an eject button next to the Benchmark model picker. Authenticated automatically with your existing API key.
  • Model browser & downloader — browse trending/popular and search HuggingFace MLX models through oMLX, and download with live progress, without leaving the app.
  • Thinking toggle (enable_thinking) for reasoning models — on the Benchmark screen, per-side in Arena, and in Flow "Set Parameters" steps.

UI & UX

  • Dark dashboard refresh: true-black canvas, accent glow, crisper cards, tighter Session Details.
  • Reports comparison now leads with prominent Fastest / Most Efficient winner cards.
  • Flow Report and the live Flow Run view are now resizable windows that remember size and position.
  • The Monitor floating HUD no longer minimizes the main window.
  • Benchmarks keep running when you switch tabs, with a live "Running … tok/s" indicator in the sidebar; every chart shows a placeholder before a run.

Fixes

  • Builds from source out of the box — no manual Secrets.swift step (#16).
  • Pins localhost connections to IPv4 (oMLX binds 127.0.0.1); clearer errors when a downloaded model isn't chat-capable.
  • Run History no longer shows empty per-core chart cards.

Download

Download anubis-oss.zip, unzip, and drag to Applications. Signed and notarized.

Requirements

  • macOS 15.0+ (Sequoia)
  • Apple Silicon (M1–M5+)
  • An inference backend (Ollama, oMLX, LM Studio, mlx-lm, …)
Assets 4
Loading