Repository navigation
Which file do I need?
The downloads are at the very bottom of this page, in the Assets list underneath all these notes. Scroll all the way down, and click Assets if the list is folded. You need exactly one file:
| Your computer | Download this ONE file |
|---|---|
| Windows 10/11 | HilbertRaum-0.1.65-portable.exe |
| macOS (Apple Silicon) | HilbertRaum-0.1.65-mac-arm64.app.zip |
| Linux | HilbertRaum-0.1.65.AppImage |
The other files are NOT the app: the llama-runtime-*.zip files carry the AI engine for offline/air-gapped installs only (normally the app offers that same download built in), SHA256SUMS.txt is for optional download verification, and the two Source code archives are the source for developers, not a runnable app.
This download is the app only — a working chat needs two more downloads, both offered on the AI Model screen inside the app: the AI engine (llama.cpp — the screen shows an install banner until it is present; without it, models run in demo mode with simulated answers) and an AI model of your choice. Two optional downloads are offered in the app only when you need them: the knowledge-pack tools, and the text-recognition (OCR) files for scanned PDFs and photos (German + English, about 4 MB; offered on the scan or photo in Documents and on the AI Model screen). Every download asks first and is SHA-256-verified. Offline installs: the llama-runtime-*.zip assets below hold the same verified engine build — unzip into the runtime/llama.cpp/ folder on the drive. Repo users can provision a drive up front instead (README step 2). Unsigned build: see the SmartScreen/Gatekeeper notes in docs/troubleshooting.md.
Fixed
- Your own llama.cpp settings on a computer no longer change how HilbertRaum runs. If you use
llama.cpp yourself and keep aconfig.inifor it, or have setLLAMA_ARG_…environment
variables, HilbertRaum's AI engines picked those up too. Such a setting could, for example, move
the chat model off the graphics card, or tell the engine to reach other computers. The engines
now start with HilbertRaum's own settings only. On Linux and macOS a system-wide
/etc/llama.cpp/config.inistill applies (#730). - The offline safeguard now notices unencrypted connections too. While HilbertRaum is offline,
it records in your Activity log when the app's main process tries to reach another computer. It
only ever saw encrypted (HTTPS) attempts, so an unencrypted one would have gone unrecorded. Nothing
in the app makes such a connection while offline; this closes a gap in the safeguard itself (#751). - A download you accepted no longer shows up as "Network attempt noticed". On an encrypted
workspace the offline safeguard decided once, when the app started, that the app was offline,
because the workspace was still locked. It then recorded every model, engine and text-recognition
download you accepted as an attempt to go online. It now checks at each connection whether the app
is offline at that moment, the way Settings → Privacy & data does. In turn, if you switch internet
access off during a session, the safeguard watches from then on. A download you started that is
still running when you lock the workspace is still recorded when it opens a new connection (#758). - Switching internet access off now stops a download that is running. The setting says the app
makes no internet calls while it is off, but a model, engine or text-recognition download you had
already started kept going. A transfer now stops the moment you switch, and a notice says why. A
file that has already arrived and is only being checked still finishes, so a model whose download
is complete still installs; if more is left to fetch after it, the download stops there, and the
notice appears then. A model download resumes where it stopped once you switch internet access
back on (#782). - You can cancel a model download during its final check. The Cancel button was greyed out for
the whole check, which takes minutes on a slow drive. It now works and stops the check at once,
and the file stays ready to pick up again. Starting a download in the moment after a cancel is
refused rather than writing to the same file twice; try again a moment later (#752). - A failed model download says what really happened. A problem on the drive now reads as one,
with no file paths in the message, and the app only promises to resume when there is something to
resume from. The Activity log records the kind of failure, not the file's path (#752). - An AI engine that fails to install no longer looks installed. If an engine archive failed its
safety check, could not be unpacked or was incomplete, its files stayed behind and the engine read
as installed. They are now removed, and the app offers the download again (#752). - A model start no longer fails with "Hashing was cancelled". Cancelling Check all model files
just as a model started could fail that start (#752). - A model download that cannot fit on the drive is refused before it starts. It used to start
and then stop partway once the drive was full, which for a large model could be well into the
download. The message now says how much the download needs and how much is free; a download you
already started counts only what is left (#768). - Sizes on a drive are shown in the same units everywhere. The download dialog showed a 14.4 GB
model as "13 GB": it counted in a larger unit than the model card and rounded down. The dialog,
the download progress, documents, saved images, knowledge packs and the text-recognition download
now all count 1 GB as a billion bytes, with one decimal, as the model card already did (#768). - The notes on the three Qwen3.8 27B files that are no longer available name their replacement.
They said a successor was still being measured; it has been in the catalog since August (#768). - The memory meter is right when no model is running. With no model loaded it measured the
conversation against the app's default window instead of the one the chosen model starts with,
usually a larger one, so the meter looked fuller than it was (#779). - No skill offer on a "couldn't find this in your documents" answer when the skill would only
repeat it. Accepting such an offer ran the same search again and gave the same answer (#779). - A list built from your documents now says that the knowledge packs you ticked were not
searched. So does a skill's answer built from a document's data. Before, such an answer showed
nothing about the packs, which read as if they had found nothing (#734). - Knowledge packs no longer stay off for the session after a rare clash at start. The program
that serves your packs picks a free connection port as it starts. If another program took that
port in the same moment, the app was meant to try once more on another one, but it did not
recognise the program's message, so packs stayed unavailable until you added or removed one or
restarted the app. It now tries again (#712). - The app's log no longer copies what the knowledge-pack program printed when it fails to
start. That message named the folder of your workspace, once for every question you asked
while packs were unavailable. The log now records only the kind of failure, as it already did
for every other knowledge-pack problem (#712). - On Windows, knowledge packs now work when your user name has an umlaut or an accent. When
the app ran from a folder such asC:\Users\Jürgen\…(a portable HilbertRaum outside a Kit keeps
its data there), no pack could be added or served, and the message blamed an incomplete file.
Packs in the app's ownzimfolder, and in other folders of that account such as your
Downloads, now work. An archive whose file name has such a character, or that sits in such a
folder elsewhere, still cannot be added; the panel now says so and how to fix it (#787). - The memory meter's colour matches the number it shows. At 74.6 % it showed "75%" in the calm
colour without the "will summarize" line, and at 89.6 % "90%" in amber. The colour and the line
now follow the number on the meter (#742). - A file dragged onto a busy chat is refused instead of lost. While an answer is being written,
or a file is still being added, the pointer now shows that the chat cannot take a drop. Before,
the file was dropped and thrown away without a word (#742). - "Send again" waits until the app knows the answer has stopped. Back on an unanswered chat
whose check kept failing, the button appeared beside an answer that might still be arriving. The
message box unlocks as before, but the button now appears only once a check succeeds. A message
that fails to send just as you press Stop now shows only its error, without a "Stopped" note as
well (#742). - A summary of every section says when it could read only some of them. Asked to summarise
each section of a document while another document, a knowledge pack or a document being
re-indexed was also in the chat, the app answered from a selection of passages without its usual
note that some sections were left out. The note is back (#736). - No document-skill offer when there is no document to work on. With documents switched off,
or only documents that never finished indexing, the chat no longer offers a document skill (such
as Meeting Minutes) under its fixed "nothing to search" reply (#733). - Keyword search in a very old workspace survives an interrupted first start. A workspace from
before keyword search existed builds its keyword index the first time a newer version opens it.
If the app was closed during that build, keyword search could stay empty for good. The build is
now one step, so an interrupted one is done again at the next start (#737). - The Performance screen clears the steps of a finished check. After Check again, the steps
of the check that had just ended, such as "Generation speed (no model running, skipped): in
progress", stayed on the card (#746). - On a Mac with Apple Silicon, switching the graphics processor off shows as off. With the GPU
switched off, or turned off after a crash, the Performance screen and Diagnostics now say the next
model start runs on the processor, instead of showing an earlier graphics measurement as current.
Checked in our tests; not yet on a real Mac (#746). - The drive preparation scripts are more careful. On Windows,
fetch-modelsno longer lets a
model past the licence check when one of its files is missing, and the scripts now work when the
drive's path contains a[.--onlywith an id that names nothing stops with a message instead
of doing nothing quietly. The Linux and macOS Kit builder now finishes on a drive that carries the
text recognition files; it used to stop at its last step (#765, #767; drive builders only;
nothing changes for an app already in use).
Changed
- The AI engine is updated to llama.cpp b12110 (the upstream v0.9.0 release). It fixes an
engine bug that, in some request patterns, made a Gemma 4 answer on a graphics card depend on an
earlier, unrelated request to the same engine; HilbertRaum's own chat requests did not run into
it in our tests (#729). In our document-question tests the built-in chat models answer as
accurately as before, and documents you already added keep their search index: it does not need
to be rebuilt. Since v0.8.0 llama.cpp also ships a program that can open the engine to other
computers over the internet; HilbertRaum never installs it. On a graphics card the Gemma 4 E2B
model reads a long question or document about 2.5 times faster in our tests. On a 24 GB graphics
card the Qwen3.8 27B models, which draft ahead to answer faster, now write about 4 % slower than
with the previous engine (their answers are the same; #800). A drive set up with the previous
engine offers Update the AI engine on the AI Model screen. - Hand-written vision model manifests must give a download source for both files or neither. A
manifest with a source for only one of them never produced a working model, even after a clean
download of several gigabytes. Such a manifest is now skipped, with a line in the log. The vision
model that ships with the app is not affected (#752). - Electron 43.8.3 → 43.8.6, the browser engine the app runs on, with the Chromium security
fixes released since 43.8.3 and Node.js 24.22.1. A download server that sends a malformed reply
can no longer make the app quit in the middle of a model, AI engine or text recognition download
(GHSA-xjj6-f33q-929c). Workspaces open exactly as before, and the minimum operating systems are
unchanged (#793). - The PDF reader is updated to pdf.js 6.5. It guards against more kinds of damaged or crafted
PDF, among them files that could make it hang or skip pages. On the 54 PDFs we compared it read
the same text as before, except one page where a few extra spaces between words are gone, and our
test scans were still recognised as before (#792). - A HilbertRaum AI Kit ships a starter set of models; the rest downloads in the app. A Kit now
comes with Qwen3 4B, the multilingual search model and Whisper small (3.2 GB) ready to use. Every
other model of the catalog is listed on the AI Model screen and downloads when you choose it. The
check that clears a Kit for sale now looks at the models it ships (#766). - The Windows drive scripts download with
curl.exeonly. It ships with every Windows from
Windows 10 version 1803 on. The scripts no longer fall back to a PowerShell download, a line a
security product had quarantined (#769; drive builders only).
Issues resolved
- #653 — GPU: an empty device-probe answer labels a GPU start cpu for the whole session
- #662 — Knowledge packs: Wikipedia's license line and German navigation boxes are delivered as article text
- #682 — Documents pipeline part 1: test-coverage gaps, and the test moves left for part 2, from the October doc-task test campaign
- #712 — Knowledge packs: a kiwix-serve start failure logs the absolute path of the workspace's transient library file
- #714 — RAG: no test proves the documents-off scope of the three registerRagIpc routing helpers
- #729 — Engine: on the pinned llama.cpp b11146 (Vulkan), Gemma 4 answers depend on earlier requests to the same server
- #730 — Engine: llama.cpp programs inherit the host's llama.cpp config files and LLAMA_ARG_* variables
- #733 — Skill offer: with documents off, the classifier offers document skills whose engine cannot run
- #734 — Knowledge packs: four deterministic answers record no pack outcome when packs are ticked
- #736 — RAG: a document mid re-index keeps its "ready" deep index for the router, so a "summarize each section" ask loses the "may miss items" heads-up
- #737 — Chat + RAG: test-coverage gaps and a suspected migration defect found in the October chat + RAG test campaign
- #738 — Docs: stale statements about chat and RAG found in the October test campaign
- #739 — Tests: leftover helper copies and test-only seams outside the chat + RAG campaign
- #742 — Chat screen: five small defects found in the chat-ui test campaign (two confirmed in code, three suspected)
- #743 — Chat screen, transcript, images and dictation: untested branches found in the chat-ui test campaign
- #744 — Docs: stale statements about the chat screen, dictation and image intake found in the chat-ui test campaign
- #745 — Tests: leftover helper copies and test-only seams outside the chat-ui test campaign
- #746 — Performance screen: finished steps stay on the idle card, and three smaller defects found in the perf test campaign
- #747 — Benchmark, Performance and picker: untested branches found in the perf test campaign
- #748 — Benchmark and Performance docs: drift found in the perf test campaign
- #749 — Tests: leftover helper copies and test-only seams outside the perf campaign
- #751 — Offline tripwire: plain-http
fetch,http.requestandnet.connectare not recorded - #752 — Downloads, models and engine install: seven small defects found in the models-drive test campaign
- #753 — Models, downloads and drive: untested branches found in the models-drive test campaign
- #754 — Models, downloads and drive docs: drift found in the models-drive test campaign
- #755 — Tests: leftover helper copies and test-only seams outside the models-drive campaign
- #758 — Offline tripwire: decided once at startup, so an accepted download reads as a violation
- #765 — build-commercial-drive.sh stops at step 7 when a drive holds two or more OCR files
- #766 — Commercial drive: the committed catalog cannot pass the sell gate
- #767 — Scripts: ten twin defects found in the models-drive test campaign, part 2
- #768 — Catalog, notices and the download dialog: eight defects found in the models-drive test campaign, part 2
- #769 — Scripts, launchers, packaging and catalog: coverage gaps left by the models-drive test campaign, part 2
- #770 — Docs and leftovers from the models-drive test campaign, part 2
- #779 — Chat + RAG: four leftovers from the #736/#737 fix (no-model meter window, empty done on a failed re-ask, offer on a no-context answer, lost keyword rows)
- #782 — Switching internet access off does not stop a running download or install
- #787 — Knowledge packs on Windows: a non-ASCII character in the drive or app-data path makes kiwix-manage and kiwix-serve fail, so packs cannot be used at all
- #790 — Sidecar spawns: a piped stdio stream with no 'error' listener can crash the main process (llama-server, GPU probe, engine check, whisper-cli)
- #795 — npm test downloads Electron mid-run on every CI leg: reranker-callbacks-composition loads the real electron unmocked (a GitHub 504 turned 4 legs red)
What's Changed
- fix(runtime): llama.cpp programs start without the host's llama.cpp config files and LLAMA_* variables (#730) by @comilionas in #731
- test(rag): pin the router's three scope counts through rag:ask (#714) by @comilionas in #732
- docs(build-state): point the #714 entry at follow-ups #733 and #734 by @comilionas in #735
- test(chat, rag): prune and repair the chat + RAG test surface (test campaign) by @comilionas in #740
- test(chat): chat-ui test campaign — chat screen, transcript, images and dictation by @comilionas in #741
- test: the performance, benchmark and picker test campaign by @comilionas in #750
- test: the models, downloads and drive test campaign, part 1 by @comilionas in #756
- Tests: the models-drive leftovers; the offline tripwire sees plain HTTP (#755, #751) by @comilionas in #757
- test: the perf campaign's leftovers; no test-only tolerance on the start, occupancy and verdict reads (#749) by @comilionas in #759
- test: the chat + RAG campaign's leftovers; required window and sendTimings, ANY_SENDER out of src/main (#739) by @comilionas in #760
- fix(performance): finished steps leave the idle card; Apple Silicon honours the GPU flags (#746, #747, #748) by @comilionas in #761
- ci: run the suite on a Mac at PR time (#753, TS-7) by @comilionas in #763
- test(models-drive): pin the #753 gaps, fix the #754 doc drift by @comilionas in #762
- fix(downloads): the models-drive campaign's seven defects (#752) by @comilionas in #764
- test: models-drive campaign, part 2 — the catalog, the shipped scripts and launchers, the notices and packaging by @comilionas in #771
- fix(scripts): build-commercial-drive.sh reaches its verdict with the committed OCR files (#765) by @comilionas in #772
- fix(kit): a Kit preloads its bundled set and lists the rest for in-app download (#766) by @comilionas in #775
- chore(deps-dev): Bump the development group with 2 updates by @dependabot[bot] in #773
- chore(deps): Bump @tanstack/react-virtual from 3.14.21 to 3.14.22 in the production group by @dependabot[bot] in #774
- fix(chat): five chat-screen defects (#742) by @comilionas in #776
- test(chat-ui): the campaign's leftovers — shared helpers, one FileDropZone, named cadences (#745, #743, #744) by @comilionas in #777
- fix(rag): a tree route the deep index cannot answer is labelled as top-k; #737's gaps pinned, the keyword index created atomically (#736, #737, #733, #738) by @comilionas in #778
- fix(scripts): the models-drive campaign's twin defects, coverage gaps and doc drift (#767, #769, #770; three of #768) by @comilionas in #780
- fix: #779 chat + RAG leftovers, #768 catalog and download-dialog defects, #743 chat-ui gaps (+ #734) by @comilionas in #781
- fix(offline-guard): the tripwire asks at every remote connect whether the app is offline (#758) by @comilionas in #783
- fix(downloads): switching internet access off stops a running download or install (#782) by @comilionas in #784
- llama.cpp b11146 → b12110 (v0.9.0); llama-connect is never installed or shipped (#729) — merge on/after 2026-11-11 by @comilionas in #785
- fix(zim): a failed kiwix-serve start is logged as its class, never with the path it printed; its bind-race retry fires again (#712) by @comilionas in #786
- docs: the non-ASCII Windows path limit points at its follow-up, #787 by @comilionas in #788
- fix(zim): Windows knowledge packs work under a non-ASCII drive or app-data root (#787) by @comilionas in #789
- fix(ci): npm test never downloads Electron; a test that loads the real electron package fails at once (#795) by @comilionas in #796
- fix(runtime): every spawn site guards its piped stdio streams, so a spawn that fails before its pipes open cannot crash the main process (#790) by @comilionas in #791
- chore(deps-dev): Bump electron from 43.8.3 to 43.8.6 by @dependabot[bot] in #793
- chore(deps): Bump the production group with 2 updates by @dependabot[bot] in #792
- test(#729): second-machine grounded-QA gate on the other five chat models (PASSED) by @humaniser in #798
- test(729): rig legs 1+2 on the RTX 3090: cache restore PASS, MTP acceptance FAIL (b12110 vs b11146) by @humaniser in #799
- release: v0.1.65 — version bump by @comilionas in #797
Full Changelog: v0.1.64...v0.1.65