OpenMindAI v5.0.0
OpenMindAI v5.0.0 is the largest release of the desktop app so far. It turns OpenMindAI from a local-only chat app into a hybrid AI workspace: fast, private on-device models by default, with an optional OpenMindAI Cloud account for hosted models — plus three new creative surfaces (Research Studio, 3D Studio, and a VS Code Coding Agent), a major image-generation and voice overhaul, and several hundred reliability, performance, and security fixes gathered since v4.0.0.
By default, v5 runs Local (on your computer — fast and private). Cloud is entirely opt-in.
Highlights — what's new since v4.0.0
Cloud + Local (hybrid AI)
- Connect an OpenMindAI Cloud account with an API key (Settings -> Cloud) to use hosted models for chat, web search, and the Studios — or stay 100% local and offline.
- Three execution modes:
- Local (default) — runs on your computer; fast and private.
- Cloud — uses hosted OpenMindAI Cloud models.
- Auto — tries Local first, falls back to Cloud when needed.
- Capability-aware routing with safe Cloud-to-Local failover, live connectivity status, and bounded reconnect/health probes.
- First-run onboarding lets you choose Cloud, Local, or Both; Cloud-only installs skip the local model download.
Research Studio (new)
- A Deep Research pipeline that plans an investigation, searches multiple sources, ranks evidence, and writes a cited manuscript.
- Advanced Web Search with multiple providers, secure configuration, and health checks.
- Citation grounding and auditing: claims are checked against the retrieved page text, near-duplicate sources are removed, and weakly-supported or mismatched numbers are flagged.
- Choose the manuscript output language (defaults to English).
3D Studio (new)
- Generate 3D scenes on your own GPU through a bundled local Python/OpenCL worker that the app starts and manages for you.
- Viewport controls, live generation status, cancel, and regenerate.
- An offline outbox saves 3D jobs when you are offline and resumes them safely, without duplicate submissions.
Coding Agent for VS Code (new)
- A token-protected Coding Agent gateway with Cloud / Local / Both routing: run tools locally while optionally using a hosted model for decisions.
- A dedicated OpenMindAI VS Code extension backed by OpenMindAI Cloud.
Image generation overhaul
- SDXL controls in the composer: seed, sampling steps, and negative prompt.
- Image renders run independently of the chat runtime, with live stage and step progress and elapsed time.
- GPU-first rendering with automatic CPU retry, the ability to stop an individual render from its card, recovery of abandoned image artifacts after a restart, and install/repair of the image renderer from Settings.
- More accurate multilingual image-intent detection (fewer wrong routes).
Voice and documents
- Kokoro text-to-speech with hash-verified model assets and non-silent synthesis checks; voice drafts in the composer.
- Document export (DOCX / PDF) with structural integrity checks, Word heading styles for outline navigation, and durable recovery of interrupted exports.
Performance
- Hardware-bounded performance profiles — Eco, Balanced, Performance, and Maximum — applied to threads, context size, batch budgets, and GPU offload.
- Context-aware response budgets and safer model context sizing based on available memory.
Chat and workspace
- Reliable Stop / cancel across model loading and streaming, with correct final status on interrupted replies.
- Long-conversation context compaction that discloses when older history is trimmed.
- Fast local full-text chat search (SQLite FTS5) with pagination and debounce.
- Paginated artifact Library and Projects lists (no more 200-item truncation), lazy project loading, and durable binary project attachments with checksums and a verified Save-a-copy flow.
- Preserved edit history, protected unsent drafts, and IME composition fixes for non-Latin input.
Bug fixes (selected)
- Fixed a first-run provisioning defect where an interrupted setup could leave a fresh install in a chat that only answered "server failed"; installation completion is now gated on the authoritative setup state reaching Completed. Legacy installs load unchanged.
- Fixed light-theme colors across the UI, plus light-mode model/dataset downloads and startup chat.
- Model readiness and health probes moved to cancellable background workers, so the UI no longer stalls while a model loads.
- More robust downloads: resumable model dependencies, concurrent-install locks, and durable package manifests that survive interrupted downloads.
- macOS secret storage now uses the native Security.framework instead of a command-line Keychain call.
- Hundreds of smaller correctness fixes across chat streaming, PDF/OCR import budgets, datasets, and exports.
Security and privacy
- Local-first by default; private image prompts are redacted from renderer failure diagnostics.
- Filesystem confinement for SQLite artifact cleanup, rendered artifacts, document exports, and portable writes.
- Hugging Face model and dependency downloads pinned to immutable revisions with verified checksums; hardened ZIP/runtime archive extraction; migration to zip 9 with a clean RustSec audit.
- Telemetry is off unless you explicitly opt in.
Also in this release
- Desktop now defaults to Local execution (fast, on-device). Cloud is opt-in via the mode toggle and does not run unless you select it.
- The chat composer shows a single mode-aware indicator: a "Cloud" badge in Cloud mode, the local model selector in Local mode, and the local selector plus a "-> Cloud" fallback hint in Auto mode (the old redundant "Cloud API model" / "OpenMindAI Cloud" dropdown has been removed).
- Research Studio defaults to English.
- Restored a green build after a dependency-group bump by pinning sha2 and sysinfo to their last API-compatible versions, aligning the iOS llama-cpp-2 pin with desktop/Android, generalizing the runtime ZIP validator, and updating SQLite reads for rusqlite 0.42.
- Version metadata synchronized to 5.0.0 across Node, Tauri, Rust, lockfiles, the portable marker, and the platform bootstrap launchers.
Upgrade from v4.0.0
You can install v5.0.0 over an existing v4.0.0 installation. Your data root, chat history, and installed local models are preserved. After upgrade, the app reports version 5.0.0, and the packaged native runtime remains present and ABI-aligned. Because the default execution mode is now Local, a fresh install starts on-device; existing installs keep whatever mode they had saved until you change it.
Release assets
- Windows x64: OpenMindAI_5.0.0_x64-setup.exe, OpenMindAI_5.0.0_x64_en-US.msi
- Windows x86: OpenMindAI_5.0.0_x86-setup.exe, OpenMindAI_5.0.0_x86_en-US.msi
- Updater metadata: OpenMindAI_5.0.0_x64-setup.exe.sig, latest.json
- SHA256SUMS.txt
- Setup launchers: OpenMindAI-Setup.bat, OpenMindAI-Setup.command, openmindai-setup.sh
- openmindai.marker
The GitHub release should remain a draft until signing, updater metadata, checksums, clean installation, and v4.0.0 -> v5.0.0 upgrade validation have passed. Native runtime ABI checks, application version checks, updater signature validation, checksums, and release-asset verification remain release gates, as do CI, Rust formatting, Clippy, Rust tests, frontend validation, CodeQL, and dependency audits.