llama.cpp Windows Manager v2.3.0

• Added daily Metrics with calendar views, flexible date selection, model/profile/runtime filters, cache-hit statistics, throughput, requests, and authenticated API/llwmctl access.
• Improved model workflows with size and parameter metadata, missing-model states, right-click actions, clearer loaded status, and compact copy-friendly endpoint reports.
• Boring but important improvements: cleaner Core, service, API, and UI boundaries; faster asynchronous screens; safer process and installer lifecycle handling; stronger repository hygiene; and 581 passing tests.

These artifacts are unsigned. Verify downloads with the matching SHA-256 companion files before running them.

llama.cpp Windows Manager v2.2.0

• Refined responsive light/dark UI, clearer Settings and launch controls, filtered runtime inventories, a compact searchable Help centre, and an updated animated product tour.
• Added launch-profile groups with retention and eviction policies, transactional multi-model loading, safer companion detection, and a unified runtime install, source-download, build, and discovery workflow.
• Expanded the authenticated llwmctl control API with group/profile operations, presentation settings, and direct-session or gateway inspection that uses the stored serving key without exposing it.
• Improved session lifecycle, gateway behavior, live metrics, download/update validation, rollback, process cleanup, accessibility, and maintainable feature boundaries, backed by a larger release-hardening test suite.
• Updated all 21 language resource contracts, including localized groups and endpoint inspection; eleven packs include translated Help and the remaining packs use complete English Help fallbacks. Upgrades preserve existing models, runtimes, profiles, settings, and application data.

These artifacts are unsigned. Verify downloads with the matching SHA-256 companion files before running them.

llama.cpp Windows Manager v2.1.0

• Sharper, higher-contrast interface with clearer navigation, actions, metrics, and consistent controls in light and dark modes.
• Runtime-aware launch settings and named profiles, including profile-aware gateway routing through the standard OpenAI-compatible model catalog.
• More reliable session lifecycle, multi-slot metrics, failure diagnostics, update staging, rollback, and process cleanup.
• Updated to .NET 10 and current SQLite/security dependencies, with stricter package checks and complete 21-language localization coverage.
• In-app upgrades from v1.x and v2.0 remain supported. Models, runtimes, settings, and app data are preserved.

These artifacts are unsigned. Verify downloads with the matching SHA-256 companion files before running them.

llama.cpp Windows Manager v2.0.0

Version 2.0 makes model launching much clearer without taking away llama.cpp's flexibility.

• Automatically discovers the selected runtime's supported launch settings and renders typed, searchable controls.
• Separates model files from editable named launch profiles and shows which profile is loaded.
• Modernizes the interface with clearer hierarchy, compact settings, responsive layouts, and improved action emphasis.
• Improves multi-slot token accounting and adds aligned trend graphs for tokens, speculative tokens, and KV cache.
• Makes loading and runtime lifecycle handling more responsive and reliable, with endpoint health, verified unloads, and diagnostic history.
• Expands release hardening, automated/WPF coverage, localization, and removes the legacy LlamaCppConsole executable alias.

Existing models, runtimes, settings, and app data are preserved during upgrade.

This release is unsigned because repository signing credentials are not configured. Verify the included SHA-256 companion files and expect a Windows SmartScreen warning.

llama.cpp Windows Manager v1.1.7

llama.cpp Windows Manager v1.1.7

v1.1.7 is a major release with multi-language support (21 languages), security hardening across process supervision, the API key lifecycle, and the model gateway, plus metrics accuracy improvements for parallel-slot workloads.

Multi-Language Support

• 21 languages now available in the sidebar language selector, covering ~96% of global users: Bulgarian (human-translated), plus 17 machine-translated languages at 87-98% coverage (Spanish, Russian, German, French, Japanese, Portuguese, Chinese, Italian, Turkish, Persian, Polish, Dutch, Vietnamese, Korean, Indonesian, Czech, Swedish). Arabic and Hindi use English fallback text pending native-script translations.
• All navigation buttons, page titles, column headers, field labels, dialog strings, status messages, tray menu items, and tooltip templates are now localised across all 530 translation keys. The language selection is persisted and applied immediately on app restart.
• English is always loaded as a fallback, so missing translations display gracefully in English rather than showing raw key names.

Security Hardening

• Custom parameter guard — User-supplied custom launch parameters can no longer override security-critical llama-server flags (--host, --port, --api-key). Attempting to do so now shows a clear error message.
• Native process cleanup — Native Windows llama-server.exe processes are now bound to a Windows Job Object with KILL_ON_JOB_CLOSE, so they are terminated automatically if the app crashes or is force-killed. Previously, native processes could become orphaned and continue running on their ports.
• API key lifecycle — Disabling API key authentication now clears the stored key from settings and triggers an OpenCode credential wipe (writes EMPTY to the provider config) instead of leaving the old key in plaintext.
• API key in process commands — The model API key is now passed via the LLAMA_API_KEY environment variable instead of a command-line argument, keeping it out of process command lines visible in Task Manager or WMI. For WSL launches, WSLENV forwarding is used to avoid embedding the key in the bash command string.
• Gateway upstream timeout — Added a 15-minute HTTP timeout on gateway upstream proxy requests to prevent thread-pool exhaustion during hung model inference.
• Null-byte rejection — WSL shell quoting (BashQuote) now rejects null bytes as a defense-in-depth measure against command truncation.
• Partial download cleanup — Cancelled or failed Hugging Face model downloads now delete their .partial files automatically, preventing multi-gigabyte orphaned files from accumulating in the models folder.

Metrics Accuracy

• Parallel-slot rate correction — The generation live rate now uses llama.cpp's reported active-generation seconds (tokens_predicted_seconds_total) instead of wall-clock time between polls. This prevents rate dilution when the model is idle between requests — a 2-slot session that generates at 35 t/s during bursts no longer shows an artificially low 13 t/s average during quiet periods.
• Newer llama.cpp support — The /slots endpoint parser now handles the current next_token object format (single object) in addition to the legacy array format, so per-slot generated-token counters are correctly populated across recent llama.cpp builds.
• Broader metric matching — Prometheus counter patterns expanded to match additional metric name variants (tokens_decoded_total, tokens_generated without _total suffix), improving compatibility with evolving upstream llama.cpp metric naming.

Localization Fixes

• Fixed navigation buttons showing raw key names (e.g. Nav.Overview) instead of translated labels on app restart with a non-English saved language.
• Fixed the language selector combo being empty on first launch after initialisation.
• Fixed active navigation highlighting breaking for non-English languages due to Loc.T() being used for internal page identifiers.
• Fixed the UpdatesNavButton label being overwritten with hardcoded English by the background update checker.
• Localised ~80 previously hardcoded English strings across page factories, view models, dialog factories, column headers, and status messages.

UI

• Added a visual separator between the language selector and the navigation buttons in the sidebar, matching the existing Tools section divider.

Build & Dependencies

• Bumped app metadata, visible UI version labels, and docs to v1.1.7.
• 19 new embedded localisation JSON resources added to the build.
• New ProcessJobObjectService infrastructure service for native Windows Job Object management.

Upgrade Notes

• Existing models, runtimes, logs, cache, state, and settings are preserved by installer update/repair and by default uninstall.
• No configuration changes are required. The language selector defaults to English; select another language from the sidebar dropdown to switch.
• This release is unsigned. Verify the .sha256 companion assets and expect Windows SmartScreen or publisher warnings until a trusted signing certificate is used.

llama.cpp Windows Manager v1.1.5

llama.cpp Windows Manager v1.1.5

v1.1.5 is a focused follow-up release for OpenCode output limits, runtime
dashboard accuracy, context checkpoint compatibility, and clearer companion
GGUF naming guidance.

Latest Changes

• Added Settings > OpenCode > Limit output and made synced OpenCode
  limit.output values follow that setting instead of each model's llama.cpp
  Max tokens launch setting.
• Renamed the OpenCode sync preference to Auto-sync entries and made the
  setting gate automatic OpenCode rewrites after Settings saves, launch-setting
  saves, variant saves, and generated API-key persistence.
• Updated OpenCode config writes so model limit blocks preserve context/output
  values intentionally and omit default output caps when no output limit should
  be written.
• Fixed context checkpoint launch arguments for current llama.cpp builds by
  using --checkpoint-min-step and no longer sending the removed
  --checkpoint-every-n-tokens flag.
• Improved Overview runtime metrics so context display uses the correct total
  for parallel slots and KV-unified mode, and so multi-slot context sizes are
  not double-counted from /slots.
• Polished the Model Status card sizing and loaded-duration handling so stale
  loading timers do not refresh or overwrite status when no load was in
  progress.
• Documented exact companion GGUF auto-detection naming rules in README and
  in-app Help for vision/projector files and upstream draft/MTP assistant files.
• Added an advanced Custom params launch setting for raw llama-server flags
  such as --n-cpu-moe, with quoted path parsing for values that contain
  spaces.
• Bumped app metadata and visible UI version labels to v1.1.5.

Upgrade Notes

• Existing models, runtimes, logs, cache, state, and settings are preserved by
  installer update/repair and by default uninstall.
• Existing OpenCode entries are rewritten only when Auto-sync entries is
  enabled or when you update/add entries from the OpenCode page.
• This release is unsigned. Verify the .sha256 companion assets and expect
  Windows SmartScreen or publisher warnings until a trusted signing certificate
  is used.

Verification

Full release gate verified locally on 2026-06-08:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File .\test-release-gate.ps1 -Runtime win-x64 -Configuration Release -IncludePublish -IncludeInstaller -InnoSetupPath "$env:LOCALAPPDATA\Programs\Inno Setup 6\ISCC.exe"

Result: Release build succeeded with zero warnings, release-hardening tests
passed (442/442), formatting was clean, no vulnerable packages were found,
the diff had no whitespace errors, and portable/installer artifact checks
passed locally.

llama.cpp Windows Manager v1.1.4

llama.cpp Windows Manager v1.1.4

v1.1.4 is an unsigned community release candidate focused on safer release
packaging, clearer model-serving controls, grouped settings navigation, better
OpenCode sync, and a more useful live Overview dashboard.

Highlights

• Added scoped LAN exposure: Local only, Gateway LAN only, Direct models LAN
  only, or Gateway + direct LAN.
• Added an auto-load gateway row on Overview so the shared endpoint, policy,
  LAN exposure, and loaded direct-session count are visible beside model
  sessions.
• Added Settings > OpenCode > Sync on launch save, plus clearer API-key
  disclosure: the app protects its saved key with Windows user protection, while
  synced OpenCode provider config stores the key in plain text because OpenCode
  must read it.
• Grouped Settings by category so cache, window, OpenCode, model, runtime,
  network, and log preferences are easier to scan.
• OpenCode model limit.output now follows each model's saved Max tokens
  launch setting when set, with a context-derived default for unlimited model
  launches.
• Added explicit Vision head choices and separate MTP head selection for
  compatible --mtp-head runtimes.
• Added Atomic TurboQuant CUDA Windows/WSL runtime package rows.
• Updated Overview metrics with compact normal and MTP token monitors, a live
  Slots card, idle-safe live token rates, normalized hardware metric separators,
  and a Model Status card that separates Loading/Loaded Model from Loading Time
  while preserving the final load duration.

Safety And Hardening

• Runtime package downloads now fail closed without expected size and SHA-256
  verification metadata or companion checksum files.
• Runtime package archives and portable app update archives are prevalidated
  before extraction to reject traversal paths, absolute paths, and unsafe tar
  entries.
• Auto-load gateway request bodies are bounded and oversized payloads return
  413 request_too_large.
• Native and WSL runtime stop paths now verify targeted shutdown more
  carefully; WSL cleanup writes diagnostic warnings when verification fails.
• Release scripts support -RequireCleanTree, and CI now verifies formatting
  plus git diff --check.
• The release gate has optional publish/installer smoke checks for hashes,
  aliases, PDB exclusion, and installer artifacts.

Upgrade Notes

• Existing models, runtimes, logs, cache, state, and settings are preserved by
  installer update/repair and by default uninstall.
• Older portable installs that still launch LlamaCppConsole.exe remain
  supported by the portable zip alias.
• This release is unsigned. Verify the .sha256 companion assets and expect
  Windows SmartScreen or publisher warnings until a trusted signing certificate
  is used.

llama.cpp Windows Manager v1.1.3

v1.1.3 is an unsigned community release candidate focused on safer release
packaging, clearer model-serving controls, better OpenCode sync, and a more
useful live Overview dashboard.

Highlights

• Added scoped LAN exposure: Local only, Gateway LAN only, Direct models LAN
  only, or Gateway + direct LAN.
• Added an auto-load gateway row on Overview so the shared endpoint, policy,
  LAN exposure, and loaded direct-session count are visible beside model
  sessions.
• Added Settings > OpenCode > Sync on launch save, plus clearer API-key
  disclosure: the app protects its saved key with Windows user protection, while
  synced OpenCode provider config stores the key in plain text because OpenCode
  must read it.
• Added explicit Vision head choices and separate MTP head selection for
  compatible --mtp-head runtimes.
• Added Atomic TurboQuant CUDA Windows/WSL runtime package rows.
• Updated Overview metrics with compact normal and MTP token monitors, a live
  Slots card, idle-safe live token rates, and normalized GPU metric separators.

Safety And Hardening

• Runtime package downloads now fail closed without expected size and SHA-256
  verification metadata or companion checksum files.
• Runtime package archives and portable app update archives are prevalidated
  before extraction to reject traversal paths, absolute paths, and unsafe tar
  entries.
• Auto-load gateway request bodies are bounded and oversized payloads return
  413 request_too_large.
• Native and WSL runtime stop paths now verify targeted shutdown more
  carefully; WSL cleanup writes diagnostic warnings when verification fails.
• Release scripts support -RequireCleanTree, and CI now verifies formatting
  plus git diff --check.
• The release gate has optional publish/installer smoke checks for hashes,
  aliases, PDB exclusion, and installer artifacts.

Upgrade Notes

• Existing models, runtimes, logs, cache, state, and settings are preserved by
  installer update/repair and by default uninstall.
• Older portable installs that still launch LlamaCppConsole.exe remain
  supported by the portable zip alias.
• This release is unsigned. Verify the .sha256 companion assets and expect
  Windows SmartScreen or publisher warnings until a trusted signing certificate
  is used.

llama.cpp Windows Manager v1.1.2

This release renames llama.cpp Console to llama.cpp Windows Manager and
turns the app into a multi-runtime, multi-model Windows manager for
llama.cpp. The normal path is now simple: install an official prebuilt
runtime, choose Windows or WSL per model, and load one or more models on stable
endpoints. Source builds are still available when you need custom or advanced
runtime work.

Highlights

• Added official prebuilt llama.cpp runtime downloads as the main workflow.
• Renamed the app to llama.cpp Windows Manager.
• Added native Windows runtime support alongside Ubuntu/WSL runtimes.
• Added a CUDA download preference in Runtimes so users can choose the
  newest CUDA package or the CUDA 12 compatibility package when upstream
  publishes both.
• Added Intel Arc GPU support through SYCL runtime choices for Windows and WSL
  when upstream packages and the local driver/tool stack support them.
• Added multi-model loading: run more than one model at the same time on
  separate saved model ports when your hardware has enough capacity.
• Added stable per-model OpenCode endpoints using separate local providers, so
  OpenCode can address concurrently served models across app sessions.
• Source downloads and builds remain available behind Runtimes > Show
  advanced.
• Windows and WSL Linux setup pages moved under Tools, since they are now
  advanced setup/troubleshooting pages rather than the normal first-run path.
• The Overview page now focuses on loaded model sessions, model size, state,
  runtime, endpoint, and live metrics for the selected model.
• Settings now keeps API key Show, Copy, and Generate actions in one
  compact action cell.

Recommended Workflow

1. Open Runtimes.
2. Install the official prebuilt runtime you want: Windows or WSL, then CPU,
   CUDA, Vulkan, or Intel Arc SYCL.
3. Open Models, download or register a GGUF model, and save its runtime and
   model port.
4. Open Overview and load one or more models.
5. Open OpenCode only if you want the app to write local model entries for
   OpenCode.

Use Tools > Windows, Tools > WSL Linux, and Runtimes > Show advanced
only for source builds, custom runtime branches, missing toolchains, or deeper
troubleshooting.

Compatibility

• Windows 10/11 x64 desktop app.
• Native Windows llama-server.exe runtimes.
• Ubuntu/WSL llama-server runtimes.
• CPU fallback plus GPU runtimes for NVIDIA CUDA, Vulkan-capable devices, and
  Intel Arc SYCL where supported.

Official prebuilt package availability depends on the upstream llama.cpp
release assets. GPU runtime success also depends on local drivers and WSL GPU
visibility.

Upgrade Notes

• Existing app data, models, runtimes, logs, cache, and settings are preserved
  by installer update/repair.
• Existing llama.cpp Console links on GitHub redirect to the renamed
  repository, and legacy portable-update paths remain supported by the v1.1.2
  zip.
• Existing models can keep using their saved launch settings; per-model ports
  are now the preferred way to keep OpenCode endpoints stable.
• If you previously built runtimes from source, you can keep them or install the
  matching official prebuilt runtime from Runtimes.

llama.cpp Console v1.1.0

This release turns the app into a multi-runtime, multi-model llama.cpp
console for Windows users. The normal path is now simple: install an official
prebuilt runtime, choose Windows or WSL per model, and load one or more models
on stable endpoints. Source builds are still available when you need custom or
advanced runtime work.

Highlights

• Added official prebuilt llama.cpp runtime downloads as the main workflow.
• Added native Windows runtime support alongside Ubuntu/WSL runtimes.
• Added Intel Arc GPU support through SYCL runtime choices for Windows and WSL
  when upstream packages and the local driver/tool stack support them.
• Added multi-model loading: run more than one model at the same time on
  separate saved model ports when your hardware has enough capacity.
• Added stable per-model OpenCode endpoints using separate local providers, so
  OpenCode can address concurrently served models across app sessions.
• Source downloads and builds remain available behind Runtimes > Show
  advanced.
• Windows and WSL Linux setup pages moved under Tools, since they are now
  advanced setup/troubleshooting pages rather than the normal first-run path.
• The Overview page now focuses on loaded model sessions, model size, state,
  runtime, endpoint, and live metrics for the selected model.

Recommended Workflow

1. Open Runtimes.
2. Install the official prebuilt runtime you want: Windows or WSL, then CPU,
   CUDA, Vulkan, or Intel Arc SYCL.
3. Open Models, download or register a GGUF model, and save its runtime and
   model port.
4. Open Overview and load one or more models.
5. Open OpenCode only if you want the app to write local model entries for
   OpenCode.

Use Tools > Windows, Tools > WSL Linux, and Runtimes > Show advanced
only for source builds, custom runtime branches, missing toolchains, or deeper
troubleshooting.

Compatibility

• Windows 10/11 x64 desktop app.
• Native Windows llama-server.exe runtimes.
• Ubuntu/WSL llama-server runtimes.
• CPU fallback plus GPU runtimes for NVIDIA CUDA, Vulkan-capable devices, and
  Intel Arc SYCL where supported.

Official prebuilt package availability depends on the upstream llama.cpp
release assets. GPU runtime success also depends on local drivers and WSL GPU
visibility.

Upgrade Notes

• Existing app data, models, runtimes, logs, cache, and settings are preserved
  by installer update/repair.
• Existing models can keep using their saved launch settings; per-model ports
  are now the preferred way to keep OpenCode endpoints stable.
• If you previously built runtimes from source, you can keep them or install the
  matching official prebuilt runtime from Runtimes.