![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4){kind=small}
Citations
Answers that name their source.
Every answer comes with clickable references back to the spot in the original document. If the model cannot back something up, it says so.
See the architecture
Local AI knowledge assistant
Query your own documents — fully offline.
LokLM keeps your documents encrypted on-device and answers questions through a chat interface — with clickable citations. No cloud, no external AI APIs.
100% offline Local encryption Open source · MIT
Our stack
Standing on the shoulders of the open-source community
Why
Local AI is hard. We make it usable.
Comparing models, building RAG pipelines, running inference locally, encrypting data — that is a lot of work before you can ask the first question. LokLM takes that part off your hands.
The hard way
- Find, evaluate, and quantize models
- Build a RAG pipeline — chunking, embeddings, retrieval, reranking
- Set up a local inference stack (llama.cpp, GPU/CPU, quant levels)
- Encrypt the vault, manage keys, plan backups
- Track current RAG research and keep up
With LokLM
Install. Drop in documents. Ask.
We handled model selection, pipeline tuning, and encryption under the hood. You do not need to be an ML engineer to query your own documents with a local AI.
How it works
In three steps.
Install, import your documents, ask — with clickable sources.
- Step 1
Drag documents into the vault
PDF, Markdown, text, or code by drag and drop. LokLM indexes locally and encrypted — no upload, no account.
- Step 2
Ask in natural language
Ask about your documents. The model runs on your machine — no request goes to the network.
- Step 3
Verify the source — click straight through
Every answer carries citations. One click opens the passage in the original document.
Vault
Encrypted on your device.
Argon2id key derivation, AES-256-GCM, and a separate key per workspace. Unlocked by your password or the 18-word recovery phrase — all of it stays local.
See the architecture
Offline
No network, no problem.
The model runs locally, the index lives locally, encryption happens locally. You can revoke network access — nothing changes.
See the diagram
Study tools
Learn from your own material.
Generate quizzes from your documents, summarise long texts, and get help while writing — all grounded in your own files, all on-device.
See the features
Translation
400+ languages, no cloud.
Translate selected text or whole documents with a local translation model (MADLAD-400). Nothing leaves your device — not even the text being translated.
See the features
More features
-
Fully offline
Models run locally. No account, no telemetry ping, no calls out to external providers.
-
Clickable citations
Every answer cites your own documents — click straight through to the original passage.
-
PDF, Word, code & more
Import PDF (including scanned, via OCR), Word (DOCX), Markdown, text, HTML, and source code — organised into workspaces.
-
Encrypted vault
Argon2id key derivation and AES-256-GCM. Your password — or the 18-word recovery phrase — unlocks a master key; each workspace is encrypted under its own key.
-
Your data stays with you
Your documents are stored encrypted on your device (in the LokLM data folder) — nothing goes to the cloud.
-
Source-available
MIT licence. Audit, fork, contribute — the full source is on GitHub.
-
Translation (400+ languages)
Translate documents and passages locally with a dedicated translation model (MADLAD-400). No cloud translator, no API.
-
Audio transcription
Transcribe audio locally with Whisper, including speaker separation — save the transcript into a workspace and query it like any document.
-
Search your codebase
Index whole repositories with code-specialised embeddings; answers cite the file and line (e.g. auth.ts:88), with folder sync.
-
Study, summarise, write
Generate quizzes and summaries from your documents and use the writing assistant — all local, all from your own sources.
Security
Where your data lives — and where it does not.
A view of the data flow. The dashed line is the network — LokLM never crosses it.
- Argon2id password hashing
- AES-256-GCM encryption
- 18-word recovery phrase
- No telemetry, no account
Built for
What LokLM is good at.
Four real examples. Honest framing: this is its strength — not open knowledge questions without context.
- Lawyer
“Where is the cap rate clause in the lease?”
→ Cited at §4.2 of Lease.pdf
- Researcher
“Summarise the methodology across these three papers.”
→ With page references for each
- Consultant
“What did the client commit to in the Q3 review?”
→ Quoted from review.docx:12
- Developer
“How is the auth middleware configured in this codebase?”
→ Cited at src/main/auth.ts:88
Download
Download
Latest release. Verify the SHA-256 checksum before installing.
Version
v0.6.6
Released
2026-07-01
Size
369 MB
windows
369 MB
macos
3.0 MB
linux
330 MB
info During setup you download one model edition — about 4–8 GB depending on your choice (Lite / Standard / Pro). A stable connection is recommended.
SHA-256 — LokLM-x64.exe
c64bed5779b7b2fad095fb4ccced1d2dd09075e2c86d09b6e41c14ef96b6bb27 System requirements
Windows 10/11, macOS, or Linux (64-bit)
RAM
12 GB RAM (16 GB for Pro)
Disk
~10 GB free disk space
FAQ
Common questions.
Is LokLM really offline?
Yes. Models and index run locally. The only network activity is the one-time model download at install and updates when you initiate them.
How big are the models, and where do they come from?
At setup you pick an edition: Lite (Qwen3.5-4B, ~3.6 GB, for integrated graphics / 12 GB RAM), Standard (Qwen3.5-4B, ~4 GB, recommended), or Pro (Qwen3.5-9B, ~7 GB). Add the embedding model (BGE-M3) and the reranker (BGE Reranker v2-M3) — ~0.9 GB together, fetched from Hugging Face. After that everything runs locally.
Can I bring my own model (GGUF)?
Yes. LokLM runs GGUF models locally through llama.cpp — drop your own GGUF files into the model directory and pick them in settings. Optionally, LokLM can use a local Ollama server instead.
Does it need a GPU?
No, everything runs on CPU too. With an NVIDIA GPU it is noticeably faster — the installer can optionally pull CUDA support (~680 MB), which accelerates both the language model and translation.
Where are my documents stored?
Encrypted on your device, in the LokLM data folder of your user account — a separate encrypted store per workspace. Encrypted with AES-256-GCM, the key derived from your password via Argon2id.
Is LokLM as smart as ChatGPT or Claude?
No. Cloud models run on orders of magnitude more hardware. LokLM is optimised for something different: privacy, citations into your own documents, and fully offline use. For open knowledge questions without context, cloud models remain better — LokLM is strong when the answer is in your own files.
How do I back up my data?
Back up the LokLM data folder — it holds the encrypted key vault and your workspaces. Since everything is encrypted, you can copy the folder to cloud storage too.
What happens if I lose my password?
You can recover the vault with your 18-word recovery phrase. Without either, the vault cannot be opened — by design.