Costs that stay flat
One setup fee and a flat maintenance fee. Your bill doesn't grow every time a user asks another question.
See pricing On-premise AI on the servers you already have
Make AI yours.
BAI installs a complete retrieval and LLM inference stack on your existing CPU servers, so your team can search, query and build on your own documents and data, without buying GPUs, hiring an AI team, or sending data outside your network.
Runs on CPU serversOpen models you chooseYour data stays in-house
Why teams choose BAI
Production AI on hardware you already own.
Most teams assume AI means GPUs, per-token bills and sending data to someone else's cloud. BAI takes a different path: start on the servers you have today, keep costs flat as usage grows, and move to GPUs only when you're ready.
Your data, your model
Documents, queries and answers stay inside your network. Use the open model you prefer, and swap it whenever a better one comes out.
Read the FAQNo GPUs required
BAI's inference engine is optimized for CPUs, so you can run AI on the servers in your data center or cloud account today. Plug it in, no specialized AI team needed.
See the benchmarksSeven input types
Your organisation's knowledge, in every format.
BAI ingests documents, images, recordings, video and web content, turning scattered information into searchable knowledge for grounded AI responses.
- Benchmarked
Video
The audio track is transcribed, and text shown on screen is read from sampled frames with OCR.
Accepts: MP4, MOV, MKV, AVI, WebM
- Benchmarked
Audio
Recordings are transcribed to text on your server with whisper.cpp.
Accepts: WAV, MP3, FLAC, OGG, M4A, AAC
- Benchmarked
PDF documents
Text is extracted page by page. Scanned pages without a usable text layer go through OCR.
Accepts: PDF
- Benchmarked
JPEG images
Text in photos, scans and screenshots is read with Tesseract OCR.
Accepts: JPEG, plus PNG, TIFF, BMP and WebP
- Supported · not yet benchmarked
Word documents
Paragraph text is extracted directly from the document, without OCR.
Accepts: DOCX, plus legacy DOC
- Supported · not yet benchmarked
PowerPoint presentations
Text is extracted from each slide, in slide order.
Accepts: PPTX
- Supported · not yet benchmarked
Web URLs
Web pages you choose are fetched and their content extracted. This is the one input that needs network access.
Accepts: http and https links
From file to answer
Extracted text is split into passages with their source and page, indexed for keyword and semantic search, and returned as sources with every answer.
Measured ingestion rates
Extraction quality and speed vary by format and by file: a clean text PDF is faster and more accurate than a low-resolution scan or a noisy recording.
How BAI gets installed on-premise
The same product. Inside your infrastructure.
Not a hosted API you rent from us, and not a custom science project. A dependable AI layer that meets your team where it is.
Scope
We look at your documents, users and servers, and size the deployment with you.
Install
BAI's services install on a server you control: Windows Server or Linux. No GPU needed.
Load models
Models are copied from local media or your own network. Answering needs no internet connection.
Ingest and ask
Your team adds documents, recordings and slides, then starts asking questions.
Prerequisite: an x86 server you control
What it takes to run an LLM on-premisePrivate LLM inference
Open-weight LLMs served on your own CPUs, tuned for your hardware.
Hybrid search over your documents
Keyword (BM25) and semantic search, combined, with the source passages returned for every answer.
A system your team owns
One secure layer for retrieval, inference and the models you choose.
Bring in what you already have
Documents, scans, recordings, slides and web pages, through one ingestion service.
Who it's for
For teams building what comes next.
Whether you are shipping your first copilot or scaling an AI-native product, BAI gives your team the control to move with confidence.
01
AI-native companies
Cut your inference bill by moving retrieval and inference onto servers you control.
02
Enterprise AI teams
Own the economics behind your product as users and usage multiply.
03
Regulated organizations
Keep sensitive workflows and data within the boundaries your customers expect.
04
Teams starting their AI journey
Start with AI on the servers you already have, then grow into GPUs when you're ready.
Measured, not estimated
Private AI performance on a CPU server.
Every figure is tied to its hardware, workload and run. Accuracy results stay unpublished until our question sets are reviewed.
- Median time to a full answer
- 9.3 s
- 1 user, 93 questions
- Median search latency
- 9.2 ms
- 1 user, top 5 passages
- Text PDF pages ingested per minute
- 124.6
- 3,236 pages
- Faster than real time for audio
- 6.4×
- 2.22 hours transcribed
Google Cloud c3-standard-22 · 22 vCPU Intel Xeon · no GPU · 29–30 October 2026
A different pricing philosophy
Own the runtime. Lose the surprise.
No per-token metering, ever. Pay a fixed setup cost to install the stack and a flat maintenance fee to keep it running—not a bill that grows every time your product gets used more.
Monthly cost, same volume
Metered cloud API$1,500/mo
BAI on your servers$400 flat
Illustrative, same volume. Real savings depend on your workload and hardware.
What's included
One price for the stack, installed on your servers and kept running for you.
Fixed setup, flat maintenance
No per-token billing
You own the runtime
Reduced rate for early partners
FAQ
Straight answers.
Models, hardware, deployment, privacy and the limits of what BAI does today.
Which file types can BAI ingest?
Video, audio, PDF, JPEG and other common image formats, Word documents, PowerPoint presentations and web URLs. Each is extracted its own way: transcription for recordings, OCR for scans and images, and direct text extraction for documents and slides. Video, audio, PDF and images have measured ingestion rates; Word, PowerPoint and web URLs are supported but not benchmarked yet.
Which models does BAI run?
Open-weight language models in GGUF format, run locally through llama.cpp. The standard configuration uses Llama 3.2 3B Instruct with 4-bit quantisation, and bge-small-en-v1.5 for search embeddings. You can change the model. Larger models need more memory and answer more slowly on a CPU.
What hardware do we need?
A server you control with an x86 CPU. No GPU is needed. Our measured test server was a 22-vCPU Intel Xeon with 86.4 GB of RAM, and the whole stack peaked at 23.7 GB of memory. We don’t publish a minimum specification yet: we size the server to your documents and number of users during scoping. Arm (Graviton) and single-GPU servers are on our benchmark roadmap.
How is BAI installed?
We scope the deployment with you, then install BAI’s services on your server: Windows Server through the BAI installer, or Linux with Docker. Models are copied onto the server from local media or your own network, so answering questions needs no internet connection. Your team then adds documents and starts asking questions.
Does our data leave our network?
Documents, questions and answers are processed on your server, and BAI doesn’t call any external AI service. The exception is web URLs: if you add one, BAI fetches that page from the internet. There is no enforced offline switch yet, so if your environment must stay fully offline, don’t add URL sources and block outbound traffic at your firewall.
What are the current security limitations?
User accounts and role-based access control aren’t built yet, and BAI doesn’t handle TLS itself. Run it inside your network, behind your existing access controls and a TLS reverse proxy. Audit logging is on the roadmap. We’re happy to walk your security team through the details.
Is every answer correct and cited?
No AI system gets every answer right. Every BAI answer returns the passages it used, with their document and page, so people can check it, but in-text citations aren’t in every answer yet. Citation correctness and answer quality are evaluated separately, and we’ll publish them once our question sets are reviewed.
Can we move to GPUs later?
Yes, that’s the intended path. llama.cpp, the inference engine BAI uses, supports GPUs, and the same models and search index carry over. BAI’s GPU configuration hasn’t been benchmarked yet; single-GPU tests are on our roadmap.
What support do you provide?
We install BAI with your team, and the flat maintenance fee covers updates and support. We agree response times with each organisation when we scope the deployment.
About BAI
Built in Nairobi, for data that can't leave the building.
BAI AI Technologies Limited is a Kenyan company based in Nairobi. We build BAI for organisations whose documents can't leave their own network, such as banks, government agencies and other regulated institutions, so they can use AI on hardware they already own.
BAI was founded by Joy Chebet, a software engineer in Nairobi. Its runtime services are written in C++ and run on standard CPU servers.
Start a conversation
Ready to make AI yours?
Tell us about your workload and we'll show you what a private, predictable AI stack could look like.