Private AI for your documents

Ask questions, get cited answers. Find amounts and deadlines. Everything runs on your Mac — nothing leaves your device.

Download on App Store See How It Works

macOS 15+ · Apple Silicon required

New in v1.4 — grounded, cited answers on every Mac, with a faster first response.

Your data stays yours. Period.

No cloud. No API keys. No telemetry. Akhna runs entirely on Apple Silicon unified memory — your documents are processed, embedded, and searched without ever leaving your Mac.

Works with WiFi off*
100% Local Processing
No Cloud, No API Keys

*After the one-time AI model download on first launch.

A knowledge engine, not a chatbot

Akhna doesn't just chat — it reads, organizes, and understands your document library.

Cited Answers

Every response links back to the exact source text. Click a citation to see where the answer came from — no hallucination, full transparency.

Calendar Integration

"When am I free Thursday?" — Akhna checks your calendar, finds free slots, and helps you plan. Toggle individual calendars on or off.

Knowledge Summary

Import documents and instantly see key facts, dollar amounts, deadlines, and people mentioned. The app understands before you ask.

Import Anything

PDF, Word, text files, and images (OCR). Drag and drop, or watch a folder for automatic imports. Scanned documents work too.

Open Original

Click any citation to open the original document at the exact page. Verify answers against your real files, not extracted text.

18 Built-In Tools

Calculate, find contacts, check deadlines, estimate reading time, convert timezones — the AI picks the right tool automatically.

Scales With Your Mac

From 3B models on 16GB to 70B on 128GB. No artificial limits — your hardware determines your experience.

How It Works

1

Import Documents

Drop PDFs and text files into Akhna, or watch a folder for automatic imports.

2

Automatic Indexing

Documents are chunked, embedded, and indexed locally using AI. Takes about a minute for a typical library.

3

Ask Anything

Type a question. Akhna searches your documents, finds the relevant passages, and streams a cited answer.

4

Verify & Explore

Click citations to open the original document. Ask about your calendar. Let Akhna find deadlines and amounts automatically.

Answers start fast — even on a 284-billion-parameter model

What matters is time to first answer — when text starts appearing — not when the last word lands. Measured on M5 Max with agent mode and full document context. Follow-up questions in the same session return the first word in under a second.

~3s
Everyday model (4B)
~7s
30B mixture-of-experts brain
~5s
284B max-quality model (ds4)
<1s
Follow-ups (any model)

Time to first word, first question of a session (~8K-token prompt); the full answer keeps streaming after. Apple's M5 Neural Accelerator speeds up the document-reading step, so the first response lands quickly even on the largest models.

Questions & Answers

What hardware do I need?

Any M-series Mac with 16GB+ RAM. But the experience varies dramatically with hardware:

Why speed depends on your Mac: AI models are memory-bandwidth bound — the GPU reads the entire model from memory for every token it generates. Faster memory = faster answers. Pro and Max chips have 2-6x the memory bandwidth of Air chips.

Why RAM matters: More RAM lets you run larger, smarter models. 16GB runs a fast 4B model. 32GB runs 7B–14B. 128GB runs a 30B mixture-of-experts brain — or ds4, a 284B model — for cloud-grade intelligence, fully private.

Mac RAM Bandwidth Best Model First answer*
M5 Max 64-128GB 400-613 GB/s 4B · 30B MoE · ds4 284B ~3-7s ⭐
M4/M5 Pro 36-48GB 200-273 GB/s 7B-14B ~5-9s
M3/M4/M5 Air 24-32GB 100-153 GB/s 4B-7B ~6-10s
M1/M2 Air 16GB 68-100 GB/s 3B-4B ~10-15s

*Time to the first word of the answer (agent routing + document search + first token); the full response keeps streaming after. Follow-up questions in the same session are near-instant thanks to warm caching. M5-family Macs lead on first-response speed because Apple's Neural Accelerator speeds up the document-reading step. Numbers are measured on M5 Max; other chips are approximate and scale with memory bandwidth.

Does my data leave my Mac?
Never. All processing — document chunking, embedding, vector search, and LLM inference — happens entirely on your device. Akhna has no server component, no analytics, and no telemetry. The only network request is to download the AI model on first launch.
What file types are supported?
PDF, Word (.docx), plain text, Markdown, and images (JPG, PNG, HEIC) via on-device OCR. PDFs with embedded text work best. Scanned documents and photos of documents are supported through Vision OCR.
How accurate are the answers?
Akhna grounds every answer in your actual document text and provides clickable citations so you can verify. It will tell you when it doesn't have enough information rather than making something up. That said, AI models can make mistakes — always check the cited sources for important decisions.
What AI models does Akhna use?
Akhna uses open-source models via MLX, optimized for each Mac's RAM. A fast 4B model is the default and runs on 16GB. High-RAM machines unlock a 30B mixture-of-experts model for deeper reasoning, and ds4 — a 284B model — for the hardest questions, all running entirely on-device. Your selected model downloads once and runs locally.
How is this different from cloud-based AI tools?
Three key differences: (1) Akhna runs 100% locally — your documents never leave your Mac. (2) Every answer cites specific passages from your documents, so you can verify claims. (3) It discovers topics and themes across your library, acting as a knowledge engine rather than just a chat interface.

Try Akhna Today

Available now on the Mac App Store. Start asking questions about your documents in minutes.

Download on Mac App Store

macOS 15+ · Apple Silicon

Want early access to new features? Join the TestFlight beta.

On Windows? Try the Akhna Explorer preview.