Datasheet, downloads and glossary
The documents a technical evaluation usually asks for, the 3D model for your own review, and plain definitions of the terms used across this site.
For the evaluation file
Specifications, installed capabilities, measured performance and pricing.
Throughput, latency and concurrent-user capacity measured on the hardware, and how the tests were run.
Each security control mapped to NIST SP 800-53 Rev. 5, NCSC CAF v4.0 and MOD Secure by Design.
The unit as a glTF binary, at real-world scale, for your own viewer or a layout review. About 1 MB.
USDZ
The same model for AR Quick Look on iPhone and iPad. About 5.5 MB.
External
RAPTOR AI LTD on the Companies House register, company number 17372642.
Background reading
- Air-gapped AIWhat it means to run AI with no network, and what an AI stack needs to do it.
- Sovereign AIKeeping models, data and infrastructure under your own control.
- SecurityThe controls on every unit, and how they map to the main frameworks.
- Pricing and licensingPublished prices, and the licence for hardware you already own.
- The App StoreThe ten AI capabilities installed on the unit, and how one is deployed.
- Frequently asked questionsModels, connectivity, concurrency, licensing and cost.
Edge AI terms, defined
The terms used across this site, in plain language. Each one can be linked to directly.
- Air-gapped AI
- Air-gapped AI is artificial intelligence that runs on hardware with no connection to the internet or to any untrusted network. Read more
- Sovereign AI
- Sovereign AI is the capability to deploy and operate artificial intelligence using infrastructure, data and model weights that remain under the legal and operational control of a single organisation or nation. Read more
- Air-gapped AI security
- Air-gapped AI security is the set of controls that keep an offline AI system trustworthy once the network is gone: what can reach the models, how updates get in, where data is stored, and what the system could send out. Read more
- Edge AI
- Running AI models on hardware close to where the data is produced, such as a vehicle, a remote site or a ship, instead of in a distant data centre. Responses are faster, and the system keeps working when the network does not.
- On-premise LLM appliance
- Hardware and software supplied together to run large language models inside your own site. Applications send requests to it over the local network instead of to a cloud API.
- Open-weight model
- A model whose trained weights are published, so it can be downloaded once and run on your own hardware. Open-weight is not the same as open source: the model’s licence still sets the terms of use.
- Pre-ingest scanning
- Checking files for malicious content before anything processes them. On an air-gapped system it covers the main remaining way in: media carried across the gap.
- Direct-attach copper (DAC) cable
- A short copper cable with a network connector fixed to each end, used to link two machines’ high-speed ports without a switch. RAPTOR’s compute nodes are joined by one, over 200Gb/s QSFP ports.
- Retrieval-augmented generation (RAG)
- Answering from your own documents. The passages most relevant to a question are found first and given to the model with it, so the answer is based on your material.
- Vision-language model
- A model that takes an image and text together and answers in text, for example describing a scene, reading a label or answering a question about a photograph.
- Time to first token (TTFT)
- How long a language model takes to start answering after it receives a prompt. People experience it as how responsive the system feels.
- Tokens per second
- How fast a language model produces text. A node’s total output is shared between the people using it, so each person’s speed falls as more people join.
- p95 latency
- The time that 95% of requests come in under. It shows the slow cases that an average hides, which are the ones a busy user notices.
Ready to deploy?
Configurations, timelines, evaluation units, or a licence for hardware you already own.