Local NotebookLM alternative: Open Notebook, SurfSense and Kotaemon on your workstation

A NotebookLM alternative that runs on your own machine

The closest self-hosted NotebookLM alternative for one person is Open Notebook: you collect sources in notebooks, get answers with citations and generate podcast-style audio overviews, with Ollama running the models on a workstation in your office, so confidential documents stay in-house. For a team, Kotaemon adds logins and shared collections. By our estimate, one RTX PRO 5000 Blackwell with 48 GB runs the text model, the search model and speech recognition together.

Google's Workspace page now calls the product Gemini Notebook and says, in its FAQ read on 30 October 2026: "Yes, NotebookLM is now Gemini Notebook." People still search for NotebookLM, and this article uses the old name for the cloud product.

The reader we have in mind is a consultant, analyst, researcher or strategy team that likes the way NotebookLM works and has client files, contracts or unpublished research that should not leave the company. The hardware sits in one section with a table, and the security settings for the IT side come after it.

What NotebookLM does, and where the files go

Google's Workspace page, read on 30 October 2026, describes the product as an assistant that "uses your uploaded sources to answer questions and fulfil requests", with answers that link back to the source, and Audio Overviews that turn documents into listenable summaries. It accepts Google Docs and Slides, PDFs, text and Markdown files, web addresses, pasted text, public YouTube links and audio files, up to 500,000 words per source or 200 MB per uploaded file. On data use the same page states: "Gemini Notebook never trains models on your uploaded data." Google's help page adds that for Workspace users, uploads, queries and responses "will not be reviewed by human reviewers", and that for other users the Google Terms of Service apply.

Whatever the terms, the sources are uploaded to Google's service to be processed. With a local tool, the sources, the search index, the chats and the generated audio stay on your hardware, and nothing goes out unless someone configures a cloud model provider. Whether a cloud service is acceptable for a given client file is a decision for your data-protection officer and the client contract; our guide on local AI and the GDPR explains what own hosting changes and what stays your duty.

Open Notebook, SurfSense and Kotaemon compared

The three projects differ most in who they are built for. The table compares them as their own pages describe them in October 2026.

GitHub repositories and licence files of lfnovo/open-notebook, MODSetter/SurfSense and Cinnamon/kotaemon, read 30 October 2026.

Open Notebook comes closest to the way NotebookLM works. It has 40.6k stars on GitHub, and its latest release appeared on 4 October 2026. It takes PDFs, Office documents, web pages, audio and video as sources, answers cite "the sources, notes and insights they used", search runs as full text and by meaning, and "transformations" produce summaries or extract insights. It offers a REST API for automation. Multi-user support appears only under "Horizon" in the project's vision document, which calls such items "directions, not dated promises".

SurfSense describes itself as an open-source NotebookLM alternative and has grown towards live data connectors for agents. Its knowledge base can sync Google Drive, OneDrive and Dropbox, and it supports local models through vLLM and Ollama. The README states that it "is actively being developed" and is "not yet production-ready", and marks team collaboration as beta. Its install script sets up Watchtower, which updates the containers daily unless you pass --no-watchtower; a company usually wants to choose when software changes. Its licence file puts most of the code under Apache 2.0 and one backend directory under the Business Source License 1.1, which allows production use as long as the software is not offered to third parties "as a commercial product or as a hosted or managed service"; your legal counsel confirms how that applies to you. Because of the beta status, the daily automatic updates and the BSL part, we do not deliver SurfSense preinstalled; for teams we set up Kotaemon.

Kotaemon is a document question-answering tool with multi-user login, private and public collections and shared chats. Its latest release appeared on 31 May 2026. It runs local models through Ollama and llama-cpp-python, reads documents with parsers such as Docling and PaddleOCR, and offers graph-based retrieval options. Its README describes no podcast feature, so it covers only the question-and-answer part of what NotebookLM does.

Open Notebook fully local: text, search and speech

A local notebook needs three kinds of model. A language model writes answers and summaries. An embedding model turns each source into vectors so that the search finds passages by meaning. A speech model reads podcast scripts aloud, and a speech-recognition model transcribes audio and video sources.

Open Notebook's guide to a fully local setup uses Ollama for the first two and names qwen3 as a general recommendation and nomic-embed-text for embeddings. For German sources we would choose a multilingual embedder such as bge-m3, a 1.2 GB download in Ollama that its page describes as supporting "more than 100 working languages", and a chat model with good German, such as Gemma 4, for which Google's model card states "out-of-the-box support for 35+ languages". Ollama sets the default context window, the amount of text the model reads at once, by the memory of the card: per its documentation page on context length, 4k tokens (a token is a word or part of a word) below 24 GiB, which is short for long sources, 32k from 24 to 48 GiB and 256k from 48 GiB; Open Notebook can override Ollama's context setting, num_ctx, which it keeps in its configuration since v1.10.0, and a larger context needs more GPU memory.

Ollama handles no speech, so voices and transcription need a second server. Open Notebook's guides for local speech both point to Speaches, an MIT-licensed server with an OpenAI-compatible interface, which Open Notebook connects to as an "OpenAI Compatible" provider. For transcription Speaches uses faster-whisper, and the guide lists Whisper models from tiny to large-v3, at about 10 GB of GPU memory for large-v3, the figure OpenAI gives for its own implementation; faster-whisper's README shows 4,525 MB for large-v2 in float16 on its test card, so we plan with 10 GB as a safe upper value. For voices the guide uses Kokoro, an 82-million-parameter model under Apache 2.0. Kokoro's voice list covers American and British English, Japanese, Mandarin, Spanish, French, Hindi, Italian and Brazilian Portuguese, and no German. Speaches also runs Piper voices, and the Piper voice collection on Hugging Face includes eight German voice folders, each with its own model card and licence: the card of de_DE-thorsten gives CC0, while de_DE-pavoque was trained on a dataset under CC BY-NC-SA 4.0, which rules out commercial use. Check the licence of each voice and listen to a sample before you use it for client-facing audio. The guide also notes that "local TTS servers usually handle one request at a time", so a long podcast takes a while. If generated podcasts go to clients, note that Article 50(2) of the AI Act requires providers of systems that generate synthetic audio to mark it in a machine-readable format, and a company that sets up such a system for its own use can count as a provider; our article on the EU AI Act for companies explains the roles, and your legal counsel confirms what applies.

You can order the workstation alone, the setup of Open Notebook, Ollama, Speaches and the models alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

Which workstation for a local NotebookLM

Everything the notebook keeps loaded must fit in the graphics card's memory. We size with our users-per-card rule from the guide to how many users one RTX PRO 6000 serves: 90 per cent of the memory the driver reports, less 3 GiB for the runtime. That gives 18.5 GiB on a 24 GB RTX PRO 4000, 40.0 GiB on a 48 GB RTX PRO 5000 and 83.0 GiB on a 96 GB RTX PRO 6000. Model sizes are Ollama's download sizes, read on 30 October 2026, and about 10 GB for Whisper large-v3 from Open Notebook's speech guide.

Our estimates by the users-per-card rule, not measurements; model sizes from ollama.com and Open Notebook's speech guide; gpt-oss-120b from our gpt-oss-120b guide.

The arithmetic behind the rows, by our estimate. On the 24 GB card, gemma4:12b (8.0 GB) and bge-m3 (1.2 GB) take 9.2 GB, or 8.6 GiB, and leave about 9.9 GiB; adding Whisper large-v3 would make 19.2 GB, or 17.9 GiB, which leaves almost nothing for context. That card suits text sources, with a smaller Whisper model for the occasional recording. On the 48 GB card, gemma4:26b (up to 19 GB), bge-m3 and Whisper large-v3 take 30.2 GB, or 28.1 GiB, and leave about 11.9 GiB for long contexts and the voice model. A team instance on one RTX PRO 6000 Workstation Edition or Max-Q runs gpt-oss-120b on vLLM, whose weights take about 65 GB (60.8 GiB) and leave room for about 19 conversations of 32,768 tokens at once, by our gpt-oss-120b guide.

The 72 GB version of the RTX PRO 5000 adds room for longer contexts or a larger model, as our comparison of the RTX PRO 5000 with 48 GB or 72 GB shows. Two RTX PRO 6000 Workstation Edition cards are the realistic maximum for that edition in one workstation; three or four Max-Q cards belong only in a chassis built for them with strong directed airflow.

We found no published speed figure for Open Notebook, SurfSense or Kotaemon on these cards. Speed depends on the card's memory bandwidth (672 GB/s on the RTX PRO 4000, 1,344 GB/s on the RTX PRO 5000, 1,792 GB/s on the RTX PRO 6000 Workstation Edition), the model size, the length of the sources and how many people ask at once. Transcribing long recordings is the heaviest step; our guide to a speech-to-text server with Whisper collects published throughput figures.

Security settings for the IT side

Open Notebook's quick-start guide states that the shipped Docker Compose file publishes the interface (port 8502) and the API (port 5055) on all network interfaces and that "there is no password by default". If other computers can reach the workstation, bind both ports to 127.0.0.1 or set OPEN_NOTEBOOK_PASSWORD, and generate your own OPEN_NOTEBOOK_ENCRYPTION_KEY instead of reusing the example. Version 1.15.0 strengthened how stored credentials are encrypted and asks for a database backup before the migration, because older versions cannot read the new format. It also sets a 180-second time-out for model calls; with slow local models, the release notes advise raising ESPERANTO_LLM_TIMEOUT and keeping it under 600 seconds.

For SurfSense, decide whether automatic daily updates fit your change process before the first start. For Kotaemon, the multi-user login separates private and public collections; who sees which collection is set by the administrator. In all three, a cloud model provider can be configured in the settings, and then the text goes there, so restrict who may change model settings.

What we install on the research workstation

When we set up the software, the workstation runs Open Notebook in Docker with Ollama on the GPU and the models already downloaded: a chat model with good German, a multilingual embedder, Speaches with Whisper for audio sources and, for podcasts, voices in the languages you need whose licence allows commercial use. The password and encryption key are set and the ports bound. For a team, Kotaemon replaces it with accounts and collections on a larger card. If a team wants one chat over shared document sets instead of notebooks, our guide to AnythingLLM as a desktop app or team server fits better, and search across SharePoint and Confluence is covered in our guide to Onyx self-hosted. For the general question of company documents in an AI tool, see our guides to preparing company documents for an AI assistant and to an AI server for consulting firms.

We deliver workstations with one RTX PRO 5000 or one RTX PRO 6000 Workstation Edition for a local notebook; send us the kind of sources you work with, the languages and how many people will use it, and we suggest the best solution for your tasks and budget.

FAQ