Skip to main content

Private AI computers

Your own AI. From one desk to the whole company.

Agents, private chat, documents and automations on hardware you control — no per-token charge, no external AI provider. Complete systems from $5,999, delivered with Y OS installed.

GB10 desk boxes · ships first

Eight chassis. One superchip.

DGX Spark, ASUS Ascent and six more — the same NVIDIA GB10 Grace Blackwell platform with 128GB unified memory in every box, each delivered with Y OS installed. Pick by chassis, storage, warranty channel and price.

Y Computer

Choose how much AI you need.

Mini is for one builder. Pro serves several assistants at once. Max puts 96GB on one CUDA GPU. Every listed system price includes Y OS, selected models, burn-in and a benchmark from your exact machine.

Compare Y Computers
Representative compact Y Computer Mini chassisPersonalMini 128One primary user who wants serious local models in a quiet desk box.128GB unified memory2TB NVMe

$5,999

complete system · Y OS installation included · sales tax and shipping extra

Learn more
Representative dual-accelerator Y Computer Pro towerTeamPro 64Studios and small teams serving several local assistants at once.256GB ECC system memory4TB NVMe

$11,999

complete system · Y OS installation included · sales tax and shipping extra

Learn more
Representative Y Computer Max 96 workstation towerMaximum compatibilityMax 96AI teams that need 96GB on one CUDA GPU and broad application support.256GB ECC system memory4TB NVMe

$24,999

complete system · Y OS installation included · sales tax and shipping extra

Learn more

Founding-batch prices include the stated hardware, build, Y OS installation, selected models, burn-in and machine-specific benchmark. Final components, warranty owner, shipping and state sales tax are confirmed in writing before shipping. Representative product images shown.

Y OS · included on every system

The private AI workspace inside every Y Computer.

Y OS turns the hardware into something people can use: private chat, cited document search, coding endpoints, voice and permissioned agents. We install, pin, test and support the complete profile.

Local by defaultEncrypted storageConnects to your existing tools (standard API)Pinned, rollback-safe updates
KnowledgeLOCAL · QWEN 3.6

Selected folders

Client research

284 files · indexed locally

Contracts

91 files · indexed locally

Product docs

418 files · indexed locally

Only folders you choose are indexed. Remove a source at any time.

Compare the renewal clauses in the three agency contracts and cite the source paragraphs.

Two agreements renew automatically after 12 months. The third moves month-to-month unless either party gives 30 days' notice.

Northwind §8.2Studio §11Relay §6.1
Ask your local knowledge…
Local inference requires no external AI provider. Web search, cloud models and connected tools are separately labeled when enabled.
Measured by YSame machine · same weights · same prompt

284B.One DGX Spark.

Y runs DeepSeek V4 Flash 0731 as a private, OpenAI-compatible endpoint at 28.29 tokens/sec 1.67× the target-only path on the exact same desk-side machine.

Flagship proof 001Passed

28.29 tok/s

fixed generation

1.67×

vs target-only

9.07 GiB

memory left

6/6

integration checks

32K

configured context

18K

longest tested input

0 KiB

process swap

The same class of AI you pay monthly for — measured on this hardware

There’s a model for everything

Text & agentsMIT
DeepSeek V4 Flash DeepSeek · 304B A13B Reads contracts, writes reports and reasons through hard problems without requiring an external AI provider. Intelligence index 50 · 103GB · 87GB on DS4 · 131K ctx
Text & agentsMIT
Ling-3.0-flash inclusionAI (Ant) · 127B MoE This week's frontier drop — big-model answers with half the box left over for everything else. IQ4_XS 66.4GB · NVFP4 81.4GB
Text & agentsApache 2.0
Muse Glimmer 30B Meta Superintelligence Lab · 30B Meta's agentic model, built for machines like this one — runs at full precision with half the box free. BF16 55.7GB · Q8_0 29.6GB
Text & agentsOpenMDW
Nemotron 3.5 Lightning NVIDIA · 30B A3B NVIDIA's brand-new agent workhorse — a million tokens of context on this exact machine. NVFP4 21.6GB · 1M ctx
CodeOpenMDW
Laguna-S-2.1 Poolside · 118B A8B A senior programmer that writes, reviews and fixes code all day without a meter running. 71GB NVFP4 · 31 tok/s
Text & agentsOpen weights
MiniMax-M2.5 MiniMax · 230B A10B The heavyweight second opinion — long documents in, considered answers out. Intelligence index 44 · 101GB · 26 tok/s
Text & agentsApache 2.0
Qwen3.5 122B Alibaba · A10B Reads a whole book in one go and answers questions about any page of it. Intelligence index 32 · 75.6GB NVFP4 · 262K ctx
Text & agentsApache 2.0
gpt-oss-120b OpenAI OpenAI's open model — fast, familiar answers for everyday questions. Intelligence index 24 · 59GB measured
Text & agentsOpen weights
Mistral Medium 3.5 Mistral AI The reliable European generalist for writing, summarising and analysis. Intelligence index 30 · ~77GB at Q4
Text & agentsOpenMDW
Nemotron 3 Super NVIDIA · 120B NVIDIA's own assistant model, tuned for exactly this hardware. Intelligence index 25 · ~61GB at NVFP4
Text & agentsOpen weights
Qwen3 Coder Next Alibaba · 80B Autocompletes, refactors and explains code at conversation speed. Intelligence index 21 · ~48GB at Q4
Text & agentsOpen weights
Devstral 2 Mistral AI Points at a real codebase and fixes bugs in it on its own. Intelligence index 19 · ~75GB at Q4
Text & agentsGemma licence
Gemma 4 26B Google · A4B The quick daily driver — snappy chat while the big models think. Q4 16.8GB · Spark build
AgentsLFM open-weight
LFM2.5-2.6B Liquid AI · 2.6B The errand-runner: clicks the buttons and calls the tools so the big model can keep thinking. ~3GB · always resident
Text & agentsOpen weights
Inkling-Small Thinking Machines A compact, sharp reasoner from Mira Murati's lab — strong logic in a small footprint. Fits with room to spare
DocumentsMIT
Unlimited-OCR Baidu · 3.3B Turns scans, PDFs and even handwriting into clean, searchable text. 6.7GB · runs beside anything
Speech-to-textMIT
Whisper large-v3 OpenAI Transcribes meetings, calls and voice notes — in almost any language. ~3GB · faster than realtime
Speech-to-textCC-BY-4.0
Parakeet NVIDIA Live transcription at conversation speed — subtitles as people speak. Realtime on-box
VoiceMIT
Chatterbox Resemble AI Clones a voice from ten seconds of audio and reads anything aloud in it. Clones from ~10s
VoiceApache 2.0
Kokoro Hexgrad · 82M Natural read-aloud voices for documents, videos and apps — tiny and instant. <1GB · instant
ImageOpen weights
FLUX.1-dev Black Forest Labs Photo-real images from a sentence — product shots, scenes, concepts. 95GB · fits, tight
ImageOpen weights
Qwen-Image-2512 Alibaba Posters and graphics with clean, readable text baked right in. 63GB · measured
ImageCommunity licence
SD 3.5 Medium Stability AI Fast image drafts — dozens of visual ideas in minutes. 22GB · 34s per image
VideoFree < $10M ARR
LTX-2.3 Lightricks · 22B Short video clips from a prompt — generated on your own desk. ~33GB working set
VideoApache 2.0
Wan 2.2 Alibaba Open video generation with the cleanest licence in the field. Fits one box
MusicOpen weights
ACE-Step v1.5 ACE Studio Full music tracks on demand — jingles, ambient beds, song demos. Full tracks · ~10GB

Scores: Artificial Analysis Intelligence Index v4.1, nine independent evals. Memory figures are measured GGUF file sizes, not estimates. DeepSeek V4 Flash measured hands-on (sources on the benchmarks page): 103GB resident, 18GB free, 131,072-token context on a single box.

Read more

What do you want to run on it?

The right Y Computer can cover all of it. Pick one.

“Runs on a single DGX Spark.”

Poolside

Laguna-S-2.1 launch, on this exact hardware class

The honest version: if nothing runs while you sleep and nothing you type is sensitive, a subscription is the better deal and we will say so. The moment agents loop — or the data is not yours to share — the answer flips. Payback maths, worked through, at why-local/cost.

Tell us what you wantyour AI to do.

One builder, a privacy-first workstation or a shared company deployment: tell us the workload, users and data boundary. We will recommend the smallest system that can do the job well.

No payment is required to get a recommendation. The exact build, lead time, warranty, shipping and tax are shown before fulfillment.