mnemosyne systems: orangu 1.4.0

mnemosyne systems is happy to announce orangu 1.4.0 - the Open Source AI-powered coding environment.

1.3.0 handed you the parts underneath the stack. 1.4.0 teaches the server to draw, makes the whole stack considerably faster, and adds two Git commands that close gaps in the everyday review workflow.

orangu remains a complete, self-contained stack that runs entirely on your own machine - a terminal environment, an on-demand model manager, and a native GGUF inference server, written end to end in Rust . No llama.cpp, no ggml, no Python. No API keys, no telemetry, and no code leaving your machine.

Image generation

orangu-server now draws pictures as well as text. Point it at an image model and every chat turn in the web console - and every /v1/chat/completions or /v1/images/generations request - is answered with a picture instead of a reply.

The recommended model is Qwen-Image 2.1, a 7-billion-parameter text-to-image and image-editing model, served from GGUF like any language model. It needs a text encoder and a vision projector beside it, and one command fetches all of them:

orangu-server download unsloth/Qwen-Image-2.1-GGUF:Q4_K_M

The same download works from Settings › Models in the web console, and Settings › Image holds the defaults every picture gets - size, steps, guidance, negative prompt, strength, and output format (PNG, JPEG, GIF, WebP, or SVG). Pictures can carry transparency, and with the vision projector loaded the model reads a picture you attach and edits it. Before you press enter, the console tells you what a picture at those settings will cost on this machine, from the server’s own measured rate.

It runs on the CPU as well as the GPU - a 512 × 512 picture is a preview in minutes on a twelve-core ARM board - and orangu-bench --image sweeps picture sizes so you can see where the time goes. The new Image generation chapter of the manual covers it end to end.

Massive performance improvements

The largest body of work in this release went into the coordinator and the server, and it touched nearly every path between a request arriving and a token leaving.

Roles share a process. Two coordinator profiles that name the same model, listener, backend, slots, and web port and differ only in role are now one orangu-server - the role travels with the request instead, so switching between coding and review no longer restarts the server
An existing server is adopted, not competed with. When the server already on the address is serving exactly what the profile asks for, the coordinator uses it as-is - no second process, no reload, no wait
The server belongs to the coordinator. A child orangu-server is stopped with the coordinator, even when the coordinator is killed outright, so no orphaned server holds on to your GPU memory
Engine work across prefill, decode, and the vocabulary projection, with orangu-bench extended to measure it honestly - fresh servers per point, warmups that settle the server’s own tuning probes before anything is timed, and head-to-head comparisons against other engines

/add_repository

Reviewing a contributor’s branch used to mean a trip to the shell to add their fork as a remote. /add_repository <user> [<branch>] derives the URL of the user’s copy from origin - on GitHub, GitLab, any other network host, or a local path - adds it as a remote, fetches the branch, and creates a local <user>/<branch> tracking it, without leaving the branch you are on:

/add_repository Jubilee101 muse
/branch Jubilee101/muse
/squash
/rebase
/create_patch

with /create_patch needed only when the rebase stops on a conflict.

/revert

/revert <commit> records a new commit that undoes an earlier one - the safe way to back out a change that has already been pushed. Tab completion offers the latest commits on the branch, newest first, a revert that stops on conflicts is resolved by /create_patch like any merge or rebase, and /revert abort abandons it.

Prometheus metrics

orangu-server serves Prometheus metrics - slot and queue gauges, latency histograms, and counters for requests and tokens - alongside a /ready probe for load balancers. A [prometheus] section binds a dedicated metrics port, so monitoring can be reached separately from the API:

[prometheus]
port = 8300

The dedicated port is a community contribution.

New model

Ternary Bonsai 2 from Prism, a 27B model with Prism’s own ternary weight types at 2.13 and 1.75 bits per weight - 7.2 and 5.9 GB for 27 billion parameters - read on the CPU, Vulkan, and Metal

Also in this release

/export issue, a PDF report of every open issue in the repository, one page per issue
/auto_review accepts glob patterns
/graph explain and /graph path - a community contribution
A session selector, and mouse selection in the terminal: drag to copy, double-click a word or a path, and double-click a tool call or reasoning block to fold it
--developer and --committer modes for orangu, and an improved /committer workflow
Logging to a file for the coordinator and the server, with log_type and log_path
Shell completions for every binary, and a better -i setup experience
Windows fixes
A streamlined README with a quick start - a community contribution
The complete manual and an updated cheatsheet ship with the release as PDF and HTML, and the manual is embedded in the binary for offline reading with /manual

Our thanks to everyone who filed an issue, opened a discussion, or sent a patch.

Upgrading

The one-liner installer puts the whole stack on the machine in a single step, on Linux, macOS, and Windows:

curl -fsSL https://mnemosyne-systems.github.io/orangu/install.sh | sh

Availability

orangu 1.4.0 is available now for Linux, macOS, and Windows on x86_64 and aarch64, with source and checksums published alongside the binaries. The backends are CPU, Vulkan, Metal, CUDA, ROCm, OpenCL, and NPU. Everything is Open Source under the GNU General Public License v3.0 .

Commercial support for orangu is available from mnemosyne systems - contact sales .

We would genuinely like your feedback - especially from anyone generating pictures with Qwen-Image 2.1 on their own hardware, and from anyone scraping orangu-server with Prometheus.

Be part of the community, and be proud of your contributions ! Try it, star it, fork it, break it, tell us what’s missing.

Thanks for your time !

Overview: https://www.mnemosyne-systems.ai/products/orangu
Repository: https://github.com/mnemosyne-systems/orangu
Release: https://github.com/mnemosyne-systems/orangu/releases/tag/1.4.0