Release history.

• GGUF: train your own model — A fifth binary, orangu-gguf, builds the model files the rest of the stack runs on. Given a manifest of permissively-licensed repositories it clones them, trains a tokenizer, packs the text into tokens, pretrains a transformer from random weights, and writes a GGUF file orangu-server loads directly — no conversion step in between. Five sizes share one vocabulary, from a tiny run that finishes in minutes to a 2b that does not.
• Quantize what you already have — The same tool rewrites an existing full-precision model at a smaller weight format: q6_k through q2_k, plus the iq4 formats for rows a K-quant’s 256-wide block will not divide. orangu-gguf --list-quantizations prints the table with a line on each.
• Workflows — jobs as a file — A YAML workflow describes jobs, their workspaces, variables, functions, explicit approvals, loops, and control flow (if, for, while, break, return, goto), over the same slash commands and prompts you type by hand. --dry-run validates every job, variable, approval, and command before anything runs, and status, pause, resume, and clear manage a run in progress.
• NPU support — A seventh backend alongside CPU, Vulkan, Metal, CUDA, ROCm, and OpenCL. Rockchip’s RKNPU is a real matmul backend on boards like the RK3588 rather than an inventory line, with the width and weight-budget rules that keep it ahead of the CPU it shares the chip with.
• /create_patch — Turns a review report into the fixes it asked for. It also resolves the conflicts a stopped /merge, /rebase, or /cherry_pick leaves behind: the model gets every unmerged path, reconstructs the intended combined behavior, and stages the result, leaving the commit or the abort to you.
• New models — Qwen 3.8-Flash-Next, including its multi-token-prediction head for drafted tokens, plus GLM 5.3 and Ornith 1.5.
• Server model list — The model list sorts, and gained a refresh-all option; LAST_USED is now tracked in plain orangu-server and not only under the coordinator. Both community contributions.
• Also in this release — Performance and reasoning work across the engine, Windows fixes, and updated dependencies.

The complete manual ships with the release as PDF and HTML, alongside the four-page cheat sheet, and is embedded in the binary for offline reading with /manual. 27 commits went into this release — our thanks to everyone who filed an issue, opened a discussion, or sent a patch.

• Eight new model families — DeepSeek 4, GLM, Kimi K3, Muse-Glimmer, Inkling-Small, Ling, NVIDIA Nemotron, and Qwen 3.8, so this year’s larger models run locally.
• KV caching — A long conversation with a large model no longer pays the full price again on every turn.
• Multi-card support — A large model can be spread across the GPUs you already have rather than the one you wish you had. Context size suggestions were raised to match.
• Model Context Protocol — orangu connects to external tools over MCP, with local server management, refresh, and reconnect handled inside orangu — no separate supervisor, and no config file to edit to bring a server back.
• Markdown rendering — A significant upgrade in the terminal, so answers arrive readable rather than as a wall of source.
• Web console: PlantUML — PlantUML renders alongside Mermaid, so class, sequence, and component diagrams come back as diagrams whichever notation the model reaches for.
• Initial license header support — A first pass at keeping license headers consistent across a codebase as you work in it.
• /developer and /committer — Two new modes: one for writing the change, one for getting it committed. Same environment, different emphasis.
• orangu-bench in the web console — A benchmark run no longer means dropping to the command line.
• Also in this release — Model last-use tracking and additional answer save formats (both community contributions), an improved /review, a cheat sheet and endpoint documentation, and macOS and WGPU stability fixes including recovery from a lost device.

The complete manual ships with the release as PDF and HTML, alongside the four-page cheat sheet, and is embedded in the binary for offline reading with /manual. 56 commits went into this release — our thanks to everyone who filed an issue, opened a discussion, or sent a patch.

• Performance: 15% to 450% faster — The attention implementation was unified across every backend, so the paths that were least optimized gained the most.
• Initial Metal support — A first Metal backend for macOS on Apple Silicon, beginning native GPU acceleration on the Mac.
• orangu-bench — A benchmarking tool, now included in the release archives, for measuring throughput on your own hardware.
• Web console: model management — List and inspect models from the browser rather than only from the command line.
• Web console: Mermaid support — Mermaid diagrams render inline in the conversation.
• Web console: history management — Delete conversations you no longer want to keep.
• Improved machine information — More accurate OS and GPU detection, to guide backend selection and troubleshooting.
• Configurable word wrap — word_wrap can now be turned on and off.

The configuration files changed in this release — run --init for orangu, orangu-coordinator, and orangu-server after upgrading.

• A complete stack — Three cooperating programs — orangu, orangu-coordinator, and orangu-server — written end to end in Rust, all under GPL v3.
• orangu-server: a native GGUF engine — GGUF loading, tokenization, the transformer forward pass, sampling, and request scheduling implemented directly in Rust. No llama.cpp, no ggml, no Python.
• Five backends — Vulkan, CUDA, OpenCL, and ROCm, plus CPU, with automated cross-verification between them.
• Model families and quantizations — Llama, Gemma, Qwen, Phi, and Mistral architectures, from F32/F16/BF16 through the K- and I-quant formats.
• OpenAI-compatible endpoints — /v1/chat/completions, /v1/completions, and /v1/embeddings, with prefix caching, persistent slots, and native workspace file operations.
• Built-in web console — Browser chat with Markdown, syntax highlighting, LaTeX, token rates, and file attachments, embedded in the binary with no build step.
• Model inventory — list, show, download, delete, prune, suggest, and system for managing local GGUF models and reading your hardware.

• A lot of token optimizations — Save a lot of tokens for your queries.
• Coordinator tool — Control multiple LLMs; auto-start when needed.
• GGUF tool — Control your LLM models.
• Deep mode for /auto_review — Do deeper reviews of your code.
• Enhance /build tool — Better incremental builds.
• New tools /statistics and /schedule — See what you are doing, and schedule your commands.

• Knowledge graph — Visualize how code is structured.
• Semantic search — Find code using natural language.
• Massive improvements to the /build tool - languages and parallel build — Build projects inside orangu.
• New tool: /shell — Execute shell commands.
• New tool: /information — Get information about your LLM.
• Active pull request report — Check the status of all active pull requests.

• Code duplication report — Eliminate code duplication in your project.
• Show your code — /show will show you Git commits.
• Review code — /review and /auto_review will now show the surrounding code.
• Compression layer — Save tokens by automatically compress them.

• Workspace tabs — work on multiple projects at once; /workspace, create workspace, and delete workspace manage tabs from the prompt. Pass -a at startup to restore all tabs from the previous run.
• Agent skills — drop a SKILL.md into your repo to teach orangu project-specific slash commands, with ghost text and Tab completion included.
• Issue & PR control — /issue adds reviewers, assignees, and labels to GitHub and GitLab issues and pull requests directly from the prompt.
• Smarter auto-review — /auto_review can now ignore files and has improved MIME detection for more accurate coverage.
• Also: /export ghost and completion, /auto_review immediate, and various bug fixes.

• AI code review — /auto_review reviews your whole branch or a single file, sorting findings into categories you can approve or reject.
• Bug bisection — /bisect drives git bisect from the prompt to find the commit that introduced a bug.
• One-line install — install.sh / install.cmd for Linux, macOS and Windows; no build step.
• Export to PDF — /export saves the console output or a review report as a polished PDF.
• Built-in manual — /manual opens the full manual, searchable and fully offline.
• Also: /get_comments, /pending, and the orangu -s shell-completion flag.

• GitLab support — drive GitLab through glab, alongside GitHub.
• Git stash — /stash push, pop, list and drop.
• Shell completions — orangu -s for bash, zsh and fish.
• Issue & PR control — /close issues and PRs, and /restart in place.
• Session tooling — /prune old sessions and highlighted /session listings.
• Also: /log <n>, /build output, opening files from /review, and /comment bodies from files.

• Review mode — the full-screen, two-pane /review of your branch.
• Pull requests — /pull_request opens a PR for the current branch.
• GitHub comments — /comment posts comments to GitHub issues.
• Usage stats — /usage reports session time and tokens.
• Virtual terminal — a wide canvas with horizontal panning for long lines.
• Also: support for unrecognized models and a customizable banner & feedback layout.

• Sessions — persistent sessions identified by UUID, with full resume.
• Auto-resume — branch-aware resumption on startup, with /session Tab completion and empty-session cleanup.

• History editing — /squash branch commits into one, and /amend the last commit message.
• Pager integration — delta for /diff and bat for /show_file when installed.

• Connection control — /connect, /disconnect and /reload.
• Markdown rendering — assistant responses are rendered with terminal styling.
• Live throughput — the Working @ X.Y t/s indicator while tokens stream.
• Also: local comment lines that stay out of the conversation.

• Initial release — an interactive terminal client for local, OpenAI-compatible LLMs, with workspace-aware tools and streaming responses.