What file types can an on-premise AI assistant read?

PDFs and office documents, scanned pages and photographs through optical character recognition, and voice recordings through transcription. A scanned manual or a photograph of a nameplate becomes a document with a text layer, so it is chunked, cited and retrieved exactly like a born-digital PDF. Native CAD files are not read; a drawing exported to PDF is.

Does an on-premise AI appliance need a GPU?

Not for a single plant. Inplant AI runs on an AMD Ryzen AI Max+ 395 with 64 GB of unified memory and no discrete graphics card, because the integrated processor and memory pool are enough for a 3-billion-parameter model to answer in about a second. Concurrency rather than corpus size is what would justify a GPU: retrieval over a hundred thousand document sections costs milliseconds, while simultaneous questions queue at the model.

What network connections does an on-premise AI appliance make?

Inbound only. People reach it with a browser on the plant network; it makes no outbound connections at all. Every network call the software makes resolves to the unit itself, and it runs in a mode where an attempt to fetch a model fails outright rather than succeeding quietly. Every model it needs is installed before the unit ships.

Inplant AI · specifications

Q: How fast can a local AI assistant answer a question about a manual?

About a second for a question whose answer already exists as a sentence in a document, and a few seconds for one that has to be synthesised. Measured in our testing lab over a corpus of machine manuals, shift-handover voice notes and nameplate photographs: repeat runs of the same lookup returned between 351 and 613 milliseconds. The range is quoted rather than an average, because the spread is real.

Which one

: Inplant Mini; Inplant Scale
People: Up to 20; A few hundredThe scope each is sold at, not a measured ceiling. Capacity has not been run at either scale, and the row will say a number when it has.
Sites: One; One or two
Throughput: 8 cores, 32 GB; Bigger silicon, 64 GB unifiedThis is what the difference in price buys. Answering is limited by memory bandwidth and indexing by cores, and Scale has more of both, so it answers quicker, indexes a first import quicker, transcribes voice notes quicker, and holds more people asking at once.
What it knows: The sameBoth appliances run the same chat model and the same embedding model, so the answers and the citations are identical. Scale is faster at producing them; it does not know more.
Buying it: Order it, we ship and invoice; We quote first, then ship
Price: $299 / month + shipping; $2,640 / month + shippingHardware, warranty and replacement are in the month on both. Month to month, no minimum term.

Compute

: Inplant Mini; Inplant Scale
Processor: AMD Ryzen™ 7 8845HS, 8 cores; AMD Ryzen™ AI Max+ 395
Memory: 32 GB; 64 GB unified
Inference: Entirely on this unitNo discrete GPU, no accelerator card. The NPU and graphics share one pool of memory, which is what lets a 3B model answer in about a second.
Chat model: qwen2.5:3b, constrained decoding
Embedding model: nomic-embed-text, 768 dimensions

Storage

: Inplant Mini; Inplant Scale
Primary: 512 GB NVMe; SanDisk SN7100 PCIe® 4.0 M.2 2280, 4 TB
Secondary: None; SanDisk SN7100 PCIe® 4.0 M.2 2280, 1 TB
Holds: Your documents, the index built from them, and the log of what was asked

Cooling and power

: Inplant Mini; Inplant Scale
Cooling: Fan, in the chassis; Noctua NF-A12x25 HS-PWM, 120 mmScale's fan is named because it was chosen; Mini uses the one its chassis ships with.
Moving parts: OneThe fan. Storage is solid state on both.
Power: 19 V DC, external supply; AC cable, C13, US/CA
Connections: Power, one Ethernet

What it reads

Everything becomes a citable chunk. Only the locator differs: a page for documents, a timestamp range for audio, the filename for a photograph, and a sheet and zone for a drawing. Rows marked shipping soon are not in the software yet.

: Inplant Mini; Inplant Scale
Documents: .pdf .docx .pptx .xlsx .html .md .txtcites as manual.pdf, p. 47
Photographs: .png .jpg .tif .bmp .webpcites as nameplate.jpg
Voice: .wav .mp3 .m4a .flac .ogg .aaccites as handover.m4a, 2:15–3:00
Video: Not supported currently
Reads from: SMB share, SharePoint, or any folderPoint it at where the documents already are. Nothing has to be moved or re-filed first, and re-reading the same directory updates in place rather than duplicating.
Scans and photographs: OCR, through the same pipelineA scanned manual or a photograph of a nameplate becomes a one-page document with a text layer, so it chunks, cites and retrieves exactly like a born-digital PDF. OCR language is set per corpus from about fifty options; it is not guessed per file.
Spoken language: Detected per recordingTranscription identifies the language itself, so a handover note in French and one in English can sit in the same corpus.
What it knows before you feed it: General engineering knowledgeThe model has read a large amount of published technical writing across many languages, so it can define a term or describe a standard with an empty index. That lane is labelled in the interface, carries no sources, and is barred from stating any specification about your equipment.
CAD drawings: .dxf shipping soonTitle block, general notes, dimension callouts and block attributes: the text on a drawing, which is what gets asked about. Cited by sheet and grid zone, the same way a manual is cited by page. Not available yet: a .dxf uploaded today is refused. A drawing exported to PDF works now. Native .dwg is under consideration.

Latency

Measured in our testing lab, over a corpus of machine manuals, shift-handover voice notes and nameplate photographs, and not on the appliance above. We measure again on your own documents once it is running, rather than asking you to trust ours.

: Inplant Mini; Inplant Scale
Social, meta: InstantCanned, or read from index metadata. No model runs.
Lookup: 350–850 msQuoted verbatim from the index, no model. Observed between 351 ms and 613 ms across repeat runs of the same question. The range is quoted rather than an average, because the spread is real.
Synthesis: 1.4 s median3B model with constrained decoding. Zero of fifteen questions over 5 s.

Choosing the model

The obvious choice was a 4B reasoning model. It lost, and the margin is the reason the appliance feels fast. Fifteen synthesis questions, same corpus, same hardware.

Model	Median	Over 5 s	Grounded	Cited	Refused correctly
4B reasoning, terse prompt	16.2 s	15/15	12/15	14/15	5/5
3B, terse prompt	3.4 s	2/15	12/15	4/15	5/5
3B, terse + schema	1.4 s	0/15	12/15	15/15	5/5

Identical grounding, twelve times faster, and the citation rate goes from 4 of 15 to 15 of 15. Asking for citations in the prompt produced them a quarter of the time; requiring a response schema with a non-empty citations array produced them every time. The model cannot return an answer without saying where it came from. That is a structural guarantee rather than a well-worded request.

Network behaviour

The question worth asking is not whether the unit is on a network. It has to be, or nobody could ask it anything. It is which direction the connections go. The appliance listens; it never dials.

: Inplant Mini; Inplant Scale
Inbound: how people reach it: A browser on your networkLaptop, tablet at the line, or phone. Nothing is installed on the client. This traffic is between the client and the unit, and does not leave your site.
Outbound: what it dials: Nothing off the unitEvery network call the software makes resolves to 127.0.0.1. The model, the index and the documents are all on the same box, so there is nowhere else for a request to go.
While reading your documents: Still nothingEnforced rather than promised: the appliance runs in a mode where any attempt to fetch a model fails outright rather than succeeding quietly. Every model it needs is already on the unit before it ships: transcription installed during the build, OCR compiled into the software itself.
Internet access: Not required, not usedAn air-gapped VLAN with no route out is a supported configuration, not a degraded one. If the site's internet goes down, nothing about the unit changes.
Firewall rule you would write: deny any → any from the unitAllow inbound on the web port from the subnets your people are on. It needs no egress rule at all.
Updates: Downloaded and applied by youWe publish software and model updates; you fetch them and upgrade the unit when it suits you. The appliance never reaches out for one itself, which is what keeps the egress rule empty.
Telemetry: NoneThe query log stays on the unit, and is readable in the admin panel.

Anything not here

If a number you need is missing, ask for it. We would rather send you a measurement than a brochure.

Request a quote Talk to engineering

Everything in the box, and everything it was measured on.

The whole thing as a PDF