Drawings, designs and supplier data
Ask questions of CAD exports, specs and supplier files. Find every drawing that references a part. Route incoming documents to the right engineer.
Live at the Dubai AI Festival · 26–27 Oct · Startup Pavilion S07
3 Spark + 1 Station slots left for Q1 2027
Dubai pricing until 30 November
Dubai AI Festival opens in
0d
0h
0m
0s
Your AI. Your data. Your walls.
Private AI that never leaves your building.
A sealed AI appliance on NVIDIA DGX or Apple silicon, pre-loaded with five products that work over your own documents, drawings and records. Built, secured and operated by Periscope. Live in 30 days.
Built on NVIDIA DGX
Apple silicon
WatchGuard Gold Partner
Designed for HIPAA · GLBA · UAE PDPL · DIFC
San Ramon, CA · Mumbai
For the data you would never upload.
Most companies have a policy that says company data stays out of public AI. Almost none can enforce it, so the most valuable work stays manual. DeepSilo is for that work.
Patient and client records
Summarise, classify and search records without a single byte leaving the practice. Audit trail of who asked what, for the compliance file.
Contracts, source code and board papers
Grounded answers from the contract set, code search across private repositories, drafting with the firm's own precedents. Nothing trains anyone else's model.
Five products, one knowledge base
Ready on day one. No ML team required.
Co-Work
A shared workspace where teams ask, draft and decide over the same knowledge base, with roles and an audit trail.
RAG Chat
Answers grounded in your documents, every answer cited to the page it came from.
DocSort
Classifies and routes incoming documents to the right team and system, with the rules you set.
Vision Chat
Questions about drawings, photos and scans, answered from the image itself.
OCR
Scanned pages become searchable text inside the silo, with sensitive fields tagged for policy.
One knowledge base
Vectors plus a knowledge graph over your documents, and domain-tuned models on the GPU. Every product reads the same base, so an answer in RAG Chat, a route in DocSort and a drawing in Vision Chat agree.
Pick the appliance
Five sizes. Hardware included. A fixed fee you can see.
Every tier ships sealed, pre-loaded with the five products, behind a WatchGuard perimeter, and operated by Periscope until your team takes over. Prices are fixed fees including hardware; operation is a monthly line.
DeepSilo Mini
AMD Ryzen AI Max+ 395 mini PC, for a team of up to 10.
$12,500
from · fixed fee, hardware included
- Memory
- 128 GB unified
- Models
- Up to 70B-class (4-bit)
- Fine-tune
- Inference only
- Speed
- Measured per model
- Operate
- From $1,500/mo
- Live in
- 21 days
DeepSilo Spark
NVIDIA DGX Spark · GB10, for a team of up to 25.
$24,500
from · fixed fee, hardware included
- Memory
- 128 GB unified · up to 1 PFLOP FP4
- Models
- Up to 200B-class; two linked ~400B
- Fine-tune
- Yes
- Speed
- Measured per model
- Operate
- From $2,500/mo
- Live in
- 30 days
DeepSilo Studio
Mac Studio · M3 Ultra · MLX, for a team of up to 50.
$29,000
from · fixed fee, hardware included
- Memory
- Up to 512 GB unified
- Models
- Up to 600B-class (4-bit)
- Fine-tune
- Inference only
- Speed
- Measured per model
- Operate
- From $2,500/mo
- Live in
- 30 days
DeepSilo Station
NVIDIA DGX Station · GB300, for a department of 100–500.
$150,000
from · fixed fee, hardware included
- Memory
- 784 GB coherent · up to 20 PFLOPS FP4
- Models
- Up to 1T-class; 7 isolated GPU instances
- Fine-tune
- Yes
- Speed
- Measured per model
- Operate
- From $6,000/mo
- Live in
- 45 days
DeepSilo Rack · Hosted
NVIDIA DGX B200/B300 or AMD Instinct, for enterprise and government, 500+.
Quote
fixed fee by design
- Memory
- By design
- Models
- Any open-weight model; training
- Fine-tune
- Yes
- Speed
- By design
- Operate
- By quote
- Live in
- 60+ days
Specs quoted from the vendors with "up to"; speeds are measured in the Periscope lab and vary by model and quantisation. Dubai AI Festival pricing applies to builds reserved by 30 November 2026. Prices are fixed-fee list prices and may change; your written sizing governs.
Not sure which? Size it in a minuteWhat runs on it
The models, by job. Real names, real sizes, the smallest box that runs each.
Open-weight models we ship today, with the memory they need at 4-bit and the licence you would be accepting. The list is refreshed monthly; the speed of your chosen model is measured in the lab before your build ships.
Mini
≈ up to 45 GB of model
70B-class dense, or 30B-class mixture-of-experts
Spark
≈ up to 115 GB · two linked ≈ 240 GB
120B mixtures, 70B fine-tuning, 400B-class when linked
Studio
≈ up to 450 GB of model
DeepSeek-class 671B mixtures at 4-bit, inference only
Station
≈ up to 650 GB of model
1T-class, seven isolated GPU instances, fine-tuning at scale
| Model | Maker | Size | ≈ Memory, 4-bit | Smallest DeepSilo | Licence | Why we pick it |
|---|---|---|---|---|---|---|
| Llama 3.3 70B | Meta | 70B dense | ≈ 44 GB | Mini (at the limit) | Llama 3.3 Community | The default workhorse for grounded answers and drafting; the broadest tooling support. |
| gpt-oss-120b | OpenAI | 117B MoE · 5.1B active | ≈ 65 GB | Spark | Apache 2.0 | Ships in MXFP4 with adjustable reasoning effort; strong instruction following. |
| Qwen3-235B-A22B | Alibaba | 235B MoE · 22B active | ≈ 140 GB | Spark ×2 or Studio | Apache 2.0 | Top of the open-weight table with a thinking mode; 119 languages. |
| Gemma 3 27B | 27B dense | ≈ 20 GB | Mini | Gemma terms | Best quality per gigabyte; 128K context; reads images. | |
| Mistral Small 3.2 24B | Mistral AI | 24B dense | ≈ 18 GB | Mini | Apache 2.0 | Fast, text and vision, good function calling for agents. |
| DeepSeek-R1 | DeepSeek | 671B MoE · 37B active | ≈ 390 GB | Studio | MIT | Open reasoning at frontier level; every expert must sit in memory. |
| DeepSeek-R1-Distill-Qwen-32B | DeepSeek | 32B dense | ≈ 23 GB | Mini | MIT | Most of R1's reasoning in a box the Mini can run. |
| Qwen3-32B | Alibaba | 32B dense | ≈ 23 GB | Mini | Apache 2.0 | Thinking and non-thinking in one model; the Spark sweet spot. |
| Llama 4 Scout | Meta | 109B MoE · 17B active | ≈ 66 GB | Spark | Llama 4 Community | Very long context for whole-contract and whole-manual questions. |
| Phi-4-reasoning-plus 14B | Microsoft | 14B dense | ≈ 11 GB | Mini | MIT | Small and strong on maths and logic; runs on the Mini with room to spare. |
| Qwen3-Coder-480B-A35B | Alibaba | 480B MoE · 35B active | ≈ 275 GB | Studio | Apache 2.0 | Agentic coding at the top of the open table; Studio or Station. |
| Qwen3-Coder-30B-A3B | Alibaba | 30B MoE · 3B active | ≈ 22 GB | Mini | Apache 2.0 | Fast agentic coding on a Mini or Spark. |
| Devstral Small 1.1 24B | Mistral AI | 24B dense | ≈ 18 GB | Mini | Apache 2.0 | Built for software-agent frameworks; strong on multi-file edits. |
| Qwen2.5-Coder-32B | Alibaba | 32B dense | ≈ 23 GB | Mini | Apache 2.0 | Reliable completion and repair across 40+ languages. |
| gpt-oss-20b | OpenAI | 21B MoE · 3.6B active | ≈ 13 GB | Mini | Apache 2.0 | Small, tool-using, fits beside a bigger model on the same box. |
| Qwen2.5-VL-72B | Alibaba | 72B dense | ≈ 46 GB | Spark | Qwen licence | Drawings, forms and charts; reads tables and locates parts on a page. |
| Qwen3-VL-30B-A3B | Alibaba | 30B MoE · 3B active | ≈ 22 GB | Mini | Apache 2.0 | Newer vision stack for long documents and video; fits the Mini. |
| Llama 3.2 90B Vision | Meta | 90B dense | ≈ 56 GB | Spark | Llama 3.2 Community | Photos and scans with the Llama tooling you already have. |
| Granite-Docling 258M | IBM | 0.26B | ≈ 1 GB | Mini | Apache 2.0 | Turns scanned pages and PDFs into structured text, tables and formulas; the OCR workhorse. |
| olmOCR 7B | Allen AI | 7B dense | ≈ 6 GB | Mini | Apache 2.0 | Document OCR tuned on real PDFs; fast on the Mini. |
| Jais family (up to 70B) | G42 · MBZUAI · Inception | 70B dense | ≈ 44 GB | Mini (at the limit) | Apache 2.0 | Arabic-first and trained in the UAE; the natural default for Gulf customers. |
| Falcon-H1 34B | TII, Abu Dhabi | 34B hybrid | ≈ 22 GB | Mini | Falcon LLM licence | Hybrid Mamba-Transformer, multilingual including Arabic; efficient on long context. |
| Falcon Arabic 7B | TII, Abu Dhabi | 7B dense | ≈ 6 GB | Mini | Falcon LLM licence | Native Arabic data with dialects; small enough for every tier. |
| ALLaM-7B | SDAIA, Saudi Arabia | 7B dense | ≈ 6 GB | Mini | ALLaM licence | Arabic and English, built for government and enterprise use in the Kingdom. |
| Qwen3-235B-A22B | Alibaba | 235B MoE · 22B active | ≈ 140 GB | Spark ×2 or Studio | Apache 2.0 | 119 languages with strong Arabic; the big-model option for mixed-language teams. |
| Qwen3-Embedding-8B | Alibaba | 8B | ≈ 6 GB | Mini | Apache 2.0 | Top of the multilingual embedding table; 100+ languages including Arabic. |
| Qwen3-Reranker-8B | Alibaba | 8B | ≈ 6 GB | Mini | Apache 2.0 | Re-orders the top hits so the answer cites the right page. |
| BGE-M3 | BAAI | 0.57B | ≈ 2 GB | Mini | MIT | Dense, sparse and multi-vector retrieval in one small model; 100+ languages. |
| bge-reranker-v2-m3 | BAAI | 0.57B | ≈ 2 GB | Mini | Apache 2.0 | The light reranker for high-volume document sets. |
| nomic-embed-text-v2-moe | Nomic | 0.5B MoE | ≈ 2 GB | Mini | Apache 2.0 | Multilingual embeddings with adjustable sizes for cheap storage. |
| Whisper large-v3 | OpenAI | 1.6B | ≈ 4 GB | Mini | MIT | Meeting and call transcription in 99 languages, Arabic included. |
| Parakeet-TDT 0.6B v3 | NVIDIA | 0.6B | ≈ 2 GB | Mini | CC-BY-4.0 | Very fast transcription with timestamps across 25 European languages. |
| Canary-1B-Flash | NVIDIA | 1B | ≈ 3 GB | Mini | CC-BY-4.0 | Transcribe and translate across English, German, French and Spanish. |
| Kokoro-82M | Hexgrad | 0.08B | ≈ 1 GB | Mini | Apache 2.0 | Natural text-to-speech for assistants and read-backs. |
| Chatterbox | Resemble AI | 0.5B | ≈ 2 GB | Mini | MIT | Expressive text-to-speech with voice control; watermarked output. |
| Llama Guard 4 12B | Meta | 12B dense | ≈ 9 GB | Mini | Llama 4 Community | Classifies prompts and answers against a hazard taxonomy; text and images. |
| Llama Prompt Guard 2 86M | Meta | 0.09B | ≈ 1 GB | Mini | Llama 4 Community | Catches prompt injection and jailbreak attempts before they reach the model. |
| Granite Guardian 3.3 8B | IBM | 8B dense | ≈ 6 GB | Mini | Apache 2.0 | Risk detection for RAG: groundedness, relevance, harmful content. |
| ShieldGemma 2 4B | 4B | ≈ 3 GB | Mini | Gemma terms | Image safety classification for Vision Chat uploads. | |
| NeMo Guardrails + Nemoguard 8B | NVIDIA | 8B dense | ≈ 6 GB | Mini | Apache 2.0 · NVIDIA Open Model | Policy rails around every product: content safety and topic control. |
Memory is an estimate: parameters × 0.56 bytes at 4-bit plus working context; mixture-of-experts models load every expert, so the total count is what matters. "Smallest DeepSilo" leaves room for the knowledge base and a second small model. Licences are summarised; read the model card before you ship. Models listed as of October 2026.
What's inside
One sealed environment, five layers, nothing leaves.
Your AI, the Periscope AI plane, compute, a sealed network and operations: all inside one environment you own.
DeepSilo appliance
Sealed
Your AI
Assistants, retrieval over your documents, agents with guardrails. Open-weight models (Llama, Mistral, Qwen, Gemma, DeepSeek; Falcon and Jais for Arabic) or your own fine-tunes.
Roles and audit trail
Who asked what, logged for the compliance file.
Periscope AI plane
Model gateway and sensitivity router · agent identity and tool policy · MCP and skill registry · signed models and data (AI-BOM) · evaluation gates · audit trail.
Signed models and data
Every model and dataset is signed and versioned before it runs.
Compute
NVIDIA DGX Spark or DGX Station, Mac Studio with M3 Ultra, or an AMD Ryzen AI Max+ mini PC for the smallest teams. Rack and hosted options for enterprises.
Air-gap option
Fully air-gapped builds for Spark, Station and Rack; updates arrive on signed media.
Sealed network
Its own segment behind a WatchGuard Firebox · network detection and response watching east-west · endpoint protection on every host · zero-trust access for people and admins, no VPN · cloud posture watched on the SaaS side.
No outbound internet
Nothing leaves at runtime; read-only connectors to the systems you name.
Operations
24/7 detection and response · monthly security and usage report · scheduled, signed model updates · build, operate, transfer with the runbooks.
Build, operate, transfer
Your team takes the keys with the runbooks, or we keep running it.
Size it in a minute
Tell us the shape of the work. We tell you the box.
Size it in a minute. This is an estimate; the written sizing follows within 24 hours.
SizingEstimate · written sizing in 24 h
- Recommended tier
- DeepSilo Spark
- Why
- team size and model class
- Fixed fee, hardware included
- from $24,500
- Operate
- from $2,500 / mo
- Live in
- 30 days
Break-even against API spendIllustrative
- Tier from the sizing
- DeepSilo Spark
- Fixed fee
- $24,500
- Monthly saving after operation
- $3,500 / mo
- Pays back in
- 7 months
Latency, privacy and predictability usually decide before the maths does. The maths is here so you can see it.
Monthly saving is API spend minus the operate fee. Hardware is yours at the end; the operate fee covers monitoring, updates and the monthly report. Taxes and shipping extra.
What happens next
From a demo to a running silo in 30 days.
What happens next, from the first demo to the day your team takes the keys.
-
Day 0
Live demo
Twenty minutes on a real appliance, your questions, your kind of documents.
-
Day 2
Sizing in writing
Tier, fixed fee, operate fee, what ships, what we need from you.
-
Day 7
Order
One signature. Hardware ordered; the sealed build starts in the lab.
-
Day 30
Live
Installed, documents indexed, five products on, first monthly report scheduled.
-
Month 2+
Operate
Monitoring, signed model updates, usage and security report every month.
-
Month 6
Transfer
Your team takes the keys with the runbooks, or we keep running it. Your choice.
Readiness check
Six questions. One honest answer about whether you need this.
-
We hold data that must not leave our network (records, drawings, contracts, code).
-
A regulator, insurer or customer asks us how we control AI use.
-
People here already use ChatGPT, Copilot or Gemini for work.
-
Our documents live in SharePoint, Google Drive or a file server we control.
-
We plan to build assistants or agents on our own data in the next year.
-
We have, or can budget, the price of a workstation for this.
Questions
What CIOs and CTOs ask before they order
Don't see your question here? Contact us and we'll answer it directly.
Does any data leave our network?
No. The appliance runs inside your network with no outbound internet at runtime. Models, documents and answers stay on the box. Updates are applied by Periscope on a schedule you approve, from signed packages.
Is it air-gapped?
It can be. The standard build is sealed: no outbound internet at runtime and read-only connectors to the systems you name. A fully air-gapped configuration with no network connection at all is available for the Spark, Station and Rack tiers; updates then arrive on signed media.
Why not Azure OpenAI or Bedrock with a private endpoint?
Often the right answer, and most of our customers run both. DeepSilo is for the data you have decided cannot leave, for models you own, and for predictable cost at volume. The router in the Periscope AI plane decides per request; you are not choosing one or the other.
Can a desktop box really run what we need?
For a team, yes: 70B-class models at useful speed, 200B-class possible, retrieval over your documents, assistants and agents. For fine-tuning at scale or hundreds of concurrent users, Station or Rack. The written sizing gives the measured answer.
How long does it take?
Mini in about three weeks, Spark and Studio in 30 days, Station in 45 days from order, hardware lead times permitting. Day one is the demo, day two the sizing in writing, day seven the order.
Who runs it?
Periscope does, until your team can. Build, operate, transfer: a monthly report, scheduled model updates, and the same security operations Periscope runs for banks and healthcare companies.
Which models?
Open-weight models you choose (Llama, Mistral, Qwen, Gemma, DeepSeek; Falcon and Jais for Arabic) or your own fine-tunes, signed and versioned. We benchmark each on your tier before it ships.
What do we own at the end?
The hardware, the models, the indexes and the runbooks. Transfer is a line in the contract, not a negotiation.
Book a live demo
Twenty minutes on a real appliance. Your questions.
Send the form and we reply within one business day with a time and, if you asked for it, the sizing in writing.
- Phone
- USA – +1 650 710 5160India – +91 9152530544
Not ready for a demo? Ask for the free Shadow AI Snapshot instead: which AI and cloud apps already hold your company data, read-only, report in 48 hours.
Request sent
Thanks, . We'll email within one business day with a time that works for your team.