Lucebox: your local AI workstation for private inference

Name: Lucebox Zero 495
Brand: Lucebox
Availability: PreOrder

Partners

Lucebox workstation, front panel with the intake grille, head-on

Intelligence you own.

The Lucebox Zero 495 is a workstation for running large open models locally. Its GPU and up to 192 GB of unified memory work on every token together, so big models answer fast. You never pay per token.

The models come installed and tuned for this hardware, with our own kernels and speculative decoding. Set it up from lucebox.com/setup and point your coding agents at it.

See price and options

What's inside.

Drag to rotate, click a partSwipe to rotate, pinch to zoom, tap a part

Loading the 3D model

Memory

128 GB or 192 GB LPDDR5X-8533, shared by CPU and integrated GPU

Bandwidth

273 GB/s

Processor

AMD Ryzen AI MAX+ PRO 495, 16 cores / 32 threads

The CPU and the integrated GPU read one shared pool, so models that do not fit in the R9700's 32 GB GDDR6 still run, with long context, or as a second model kept loaded.
Card

AMD Radeon AI PRO R9700

Bandwidth

640 GB/s

Compute

64 compute units, up to 191 TFLOPS at FP16

The hot path runs here: the weights and KV cache of a 27B-class model stay on the card, where the custom kernels and speculative decoding run.
Capacity

2 TB, or 4 TB as an option

Interface

NVMe 1.4, M.2

It holds the system, the engine and the tuned model profiles, with room for your own weights, and fast reads mean big model files load quickly.
Ports

2× USB4, 2× USB-A 3.2 Gen 2, 2× USB-A 2.0, HDMI 2.1, DisplayPort 1.4, 10 GbE Ethernet, plus 4 display outputs on the GPU

Wireless

Wi-Fi 7 and Bluetooth 5.4, built in

Clustering

Optional cluster kit per machine: Mellanox ConnectX-5 Ex, 2× 100 GbE, on a PCIe 4.0 x4 link, installed, cable included

Most people keep it out of earshot and reach it over the network; the rear ports also drive screens when you want them.
Output

1,000 W, SFX form factor

Draw

~500 W under full inference load, ~40 W idle

One unit feeds the board and the GPU with headroom to spare, and it plugs into a standard wall outlet.
Intake

120 mm Arctic P12 Slim behind the side grille, blowing on the R9700

Compute board

Two 80 mm fans on its cooler, out through the rear duct

GPU

R9700 blower, out of the card's bracket

The side fan feeds the R9700 fresh air, and the board and the GPU each push their own heat out the back, so speed holds through long runs.
Dimensions

340 × 110 × 320 mm (13.4 × 4.3 × 12.6 in)

Weight

6 kg (13.2 lb)

Finish

Anodized black

A rigid aluminium frame with carry handles holds a full-length workstation GPU, the compute board and the power supply in a box that fits a shelf.
Covers

Parts and labor, after 72 hours of burn-in

Your data stays private
Plug-and-play setup
Runs big models fast
No per-token bill

Check full specs

What developers say.

0xSero@0xSero

Best deal without a doubt right now grab it before it’s too late. With this you’ll be able to run GLM-Flash/DS4.1-Flash really fast ❤️

Ahmad@TheAhmadOsman

I like your stuff so far, keep going

Carlo@Italianclownz

This is probably one of the best prices to get into powerful local AI right now. The Lucebox team is dedicated to support and giving you a way to run quality AI in-house for your needs. Besides being affordable for the compute power you get there is a big community involvement

Lotto@LottoLabs

This is very compelling Especially with strata type offloading I’ve been working with midrange bandwidth cards like the P150 and you can get really decent performance I expect small to mid size moes to rip on this thing

Loktar@loktar00

I'm tapped out too many 3090s and Christmas is coming but this is a HECK OF A DEAL.

Digital Spaceport@gospaceport

Given the "new" Spark size/price news, this looks like a very serious competitor. Especially with it's R9700 Pro included.

vitalik.eth@VitalikButerin

github.com/Luce-Org/lucebox-hub looks promising as a way to run "dense" models (eg. Qwen 27B) more efficiently. It's janky, but on my 5090 laptop it seems to be ~2x more tok/s than llama.cpp

Jaro@jaro3th

Best deal on any kind of hardware rn Honestly surprised that they priced it this low Congrats @pupposandro on the launch @luceboxai is the way

Marco Franzon@mfranz_on

Local AI lovers this one is for you. One of the best AMD setups for hosting your own AI models

Sudo su@sudoingX

this guy just cracked 134 tok/s on qwen 3.5-27b dense and 73 on new qwen 3.6-27b on a single 3090. open source moves at godspeed in 2026.

Donato Capitella@dcapitella

The Lucebox guys have been cooking!

Our open source inference engine is used by engineers at

60+Contributors

790+Pull requests

140K+Model downloads

Join the community

Built for speed.

DeepSeek V4 FlashMixture of experts

Decode, tokens per second

Lucebox 54
Mac M5 Max 39.35
DGX Spark 35.3

DeepSeek V4.1 Flash, 32K prompt

Lucebox ~33
Ryzen AI MAX+ alone ~16

DGX Spark and Mac M5 Max: public runs. Ryzen AI MAX+ alone: Framework's run.

Lucebox Mix

One endpoint, routed by model id

Radeon AI PRO R9700

Qwen3.8-27B

Serving 8 users

Parse a CSV file

def parse(path):
    with open(path) as f:
        return [row.split(",") for row in f]

Ryzen AI MAX+ PRO 495

DeepSeek V4 Flash

Serving 4 users

Why does this test fail at random?

Two workers write the same cache file at once. Give each worker its own file, or take a lock before writing.

1.5× DGX Spark decode speed.

Two processors work on every token. The R9700 runs the part every token needs from its own 640 GB/s memory, and 192 GB of unified memory holds the rest of the model.

On DeepSeek V4 Flash it writes faster than DGX Spark and a Mac M5 Max. On V4.1 it writes twice as fast as the same Ryzen AI MAX+ on its own.

Decode is how fast the answer comes out once the model has read the prompt. DGX Spark and Mac M5 Max are public results, each with its own setup, quantization and prompt. The Ryzen AI MAX+ alone is Framework's published run on the same chip, with its own engine.

DeepSeek V4 Flash details DeepSeek V4.1 Flash details

One model across both chips.

DeepSeek V4 Flash has 256 experts in each layer and uses 4 of them for each token. The R9700 keeps the most-used experts beside the layers every token needs, and unified memory holds the rest.

Both chips read their part at the same time, and only small results move between them. The cache stays on the GPU.

How the engine splits the work

Two models at once.

Lucebox Mix runs one model on each chip behind one endpoint: Qwen3.8-27B on the R9700, DeepSeek V4 Flash in unified memory. Your apps pick a model by its id, and the engine sends the request to that model's chip.

Each model has its own memory and its own compute. In our test, with 8 users on Qwen and 4 on DeepSeek, Qwen kept 88% of its solo speed and DeepSeek 99%.

Two models at once

Built in partnership with

See all results

Use cases.

CI/CD and code review

Builds, tests and AI code review on one box in the office, instead of runners and an API billed by the minute.

What one box replaces Linux runner minutes and hosted AI, a month $1,559 a month

Written off over the months the box serves 36 months

Pays for itself in: about 5 months
In the budget: $256 a month
Over the write-off: saves about $46,900
Then saves: $1,530 a month

Coding agents

Give agents the long jobs, like a refactor or a test suite, and run several at once with no rate limit.

connect  claude code → lucebox.localmodel    qwen3.8-27b> write tests for the upload retry logic

Private knowledge base

Ask across your wiki, specs, tickets and code, and get answers that cite the file. Nothing leaves the box.

> where do we set the upload retry limit?  maxRetries = 5, services/upload/config.ts  source: docs/runbooks/uploads.md

Always-on agents

Agents triage issues and watch alerts through the night, then leave a summary for the morning. The night costs only power.

07:00  overnight summary       9 issues triaged, 2 marked urgent       disk on db-2 at 91%, ticket opened

See the system

Easy setup, with apps you already use.

lucebox.com/setup Lucebox Manage Open setup

Loading the real setup UI…

Setup

Manage

Start at lucebox.com/setup. Pair the machine nearby and put it on your network from the setup page.

Works with the tools you already use

Claude Code

Codex

OpenCode

Hermes

OpenClaw

Oh My Pi

Open WebUI

See the platform

Order your Lucebox.

Developer

For individuals

Engine open source, free forever

Starting at $5,999 USD per machine / taxes, duties and shipping included Taxes and shipping included

Lucebox Engine, open source, with tuned models preinstalled
1-year warranty on parts and labor
Worldwide shipping included
All taxes and duties included

Order now

Card or Shop Pay Full refund within 30 days Terms

Configure yours

Open the configurator

Unified memoryLPDDR5X-8533, 273 GB/s

StorageNVMe SSD

ClusteringOptional

100 GbE +$799 per machine Cluster kit: Mellanox ConnectX-5 Ex, 2× 100 GbE, on a PCIe 4.0 x4 link, installed, cable included

Business

For teams and companies: any number of machines, installed on your network, your models tuned for them, and direct support from our engineers.

Book a call Business details (opens in a new tab)

We share what we learn.

A Lucebox with a long strip of printer paper running out of it, rows of code with a check mark every few rows

AMD · CI

CI builds and AI code review on one Lucebox: paid back in about 5 months

Our CI bots cost us about $1,340 a month on hosted models. One Lucebox reviews every pull request on our main repository and runs Linux jobs for $8,199 once and about $1 a day, and pays for itself in about 5 months.

October 2026 4 min read

AMD · TEXT-TO-IMAGE

MOSS-Image on the R9700: a 1024x1024 image in 2.7 s at 8 steps, 1.8x faster than ComfyUI on a DGX Spark

Our engine now runs MOSS-Image, a 5B text-to-image model, natively on the Radeon AI PRO R9700. A 1024x1024 image takes 2.7 s at 8 steps, 1.8x faster than ComfyUI on a DGX Spark at the same settings. The engine is open source.

October 2026 8 min read

PFlash v2: a long document with a few whole units highlighted in gold, flowing into a short compressed prompt

LONG CONTEXT · RAG

PFlash v2: 3x faster processing on long prompts

PFlash decides which parts of a long prompt the model reads: 92 of 124 public long-context questions against 94 for full prefill, with the model reading a median 27% of the prompt and the first token 3.7x sooner on average on one AMD Radeon AI PRO R9700.

October 2026 4 min read

Read all articles

Frequently asked questions

What is Lucebox?

A plug-and-play workstation for local AI inference. It runs headless on Ubuntu Server: set it up from your laptop or phone, manage it from the browser, and point your tools at it over the network. It arrives with our open source inference engine, tuned models, custom kernels, and speculative decoding ready to use through a single CLI. What a local inference PC is

What does it cost?

The Lucebox Zero 495 starts at $5,999 USD per machine for the Developer tier (individuals), with 128GB of memory and 2TB of storage. Business, for teams and companies, is priced on a call, case by case. 192GB of memory adds $2,200, 4TB of storage adds $250, and the cluster kit is $799 per machine. Import duties and worldwide shipping are included, and so are taxes outside the United States. In the US, no sales tax is added at checkout on shop.lucebox.com. Nothing is due on delivery.

How do I pay?

On shop.lucebox.com, by card or Shop Pay. The shop collects the payment and your shipping address, then emails the receipt and order number and confirms what happens next. The price includes import duties and shipping, and taxes outside the United States. In the US, no sales tax is added at checkout. Nothing is due on delivery.

What hardware is inside?

An AMD Ryzen AI MAX+ PRO 495 with 128GB or 192GB of LPDDR5X-8533 unified memory, an AMD Radeon AI PRO R9700 with 32GB of GDDR6, a 2TB or 4TB NVMe 1.4 SSD, and a 1000W 80+ Gold SFX power supply. The extruded aluminium chassis, anodized black, has a volume of 11.97 litres.

When does it ship?

Deliveries are expected in January 2027 for 128 GB machines and in February 2027 for 192 GB machines. The shop emails your receipt right away and we email tracking the day your machine leaves.

Can I order multiple units?

Yes. You can take up to 5 machines in one checkout. For more than 5, book a call and we prepare a single quote and invoice for the team. Book a Lucebox call

Book a call with us

What is the cluster kit?

A Mellanox ConnectX-5 Ex networking card with two 100 GbE ports on a PCIe 4.0 x4 link, installed and configured in your Lucebox before shipping, plus the cable. When two Luceboxes split one model, every step sends many small messages between them, so what matters most is how fast each message arrives, not peak bandwidth. RDMA writes each one straight into the other box's memory in microseconds. It is sold per machine, $799 each, and you can add it to a single Lucebox. To cluster two Luceboxes, add the kit to both, connect the cable and they work together with nothing to install. One kit also links a Lucebox to another machine with a 100 GbE port.

Is the software open source?

Yes. Our inference work is public at github.com/Luce-Org/lucebox, with 55+ contributors and 650+ pull requests. It includes hand-tuned kernels, DFlash speculative decoding, PFlash prefill, and hybrid-model optimizations.

Can I run other inference engines?

Yes. Lucebox runs Ubuntu Server with AMD ROCm and gives you full root access, so you can install other open source inference engines as on any Linux machine. The Lucebox engine is the one that comes installed and tuned for this hardware.

What about Ollama and LM Studio?

Lucebox does not require either one. It exposes an OpenAI-compatible endpoint, so tools that accept a custom base URL can connect directly while Ollama and LM Studio remain available for any separate workflows you already use.

Can I use it away from home?

Yes. During setup you sign in to Tailscale from a link and approve Lucebox, or add it later from Manage. With Tailscale on your laptop or phone, signed in to the same account, you reach Manage and the private API over an encrypted tunnel from any network, at home, in the office or on the road, without opening a router port or exposing anything to the public internet. Point Claude Code, Codex, OpenCode or any compatible client at the tailnet address and keep working. Skip it at setup and everything still works on your home network.

What's the power draw?

Around 500 watts under full inference load, with up to 300 W for the Radeon AI PRO R9700 and the remaining system power allocated to the Ryzen platform and components. Idle consumption is approximately 40 W. Plug it into any standard wall outlet.

Can I SSH in?

Yes. Full root access. It is your machine. Open the shell, install whatever you want, run whatever you want.

Do you ship internationally?

Yes. We ship worldwide, and shipping is included in the price. VAT, import duties, taxes and customs fees are included too: we pay them, so there is nothing to pay on delivery. In the United States, no sales tax is added at checkout. If a carrier ever asks you for a duty or tax payment, write to support@lucebox.com before paying. The shop collects your shipping address at checkout.

Can several people use it at the same time?

Yes. Lucebox serves concurrent requests from several users, or from one user running several agents. Partial tensor parallelism and attention paging keep throughput up as users join. Memory is the first limit: if everyone runs 256K to 1M contexts it gets tight, and we work out a setup for that case with you.

How does it compare with the DGX Spark for several users?

On paper the DGX Spark has more TFLOPs. Lucebox's custom speculative decoding models and server-side tensor parallelism put it ahead in delivered speed, for one user and for several. The engine page collects the published comparisons.

What is Business?

Lucebox for teams and companies. We install as many machines as you need on your network, tune your own models and fine-tunes for them, and set up clusters. Your team gets direct chat with our engineers, with a support response time written into the contract. Price and invoicing are agreed on a call, case by case. Business details

What models can I run?

Far more than fits in 32GB of VRAM: the engine runs one model across the R9700 and up to 192GB of unified memory, so the 284B DeepSeek V4 Flash fits. The engine ships tuned for Qwen3.8-27B and DeepSeek V4 Flash, and also runs Laguna XS 2.1, Gemma 4 and Ling 3.0 Flash with its drafters. You can bring your own GGUF; the CLI loads it. See the models

Can I use it with Claude Code, Cursor, and OpenCode?

Yes. Lucebox exposes OpenAI-compatible and Anthropic-compatible endpoints. Point a supported client or agent harness to the Lucebox base URL and keep using its normal workflow.

Does it need internet?

No. Lucebox runs fully offline once your model is loaded. Internet is only needed for updates and the optional cloud fallback.

How loud is it?

Quiet at idle; the Radeon AI PRO R9700 cooling becomes audible during sustained inference. Most people keep it where fan noise does not matter and reach it over the network, so it never has to sit on your desk. The chassis directs hot air out the back.

What's the warranty?

1 year on the full machine, parts and labor. We replace or repair any unit that fails under normal use.

What about cancellations, returns and refunds?

Within 30 days of your order you can cancel for a full refund; after that, until it ships, we refund the price minus 25%, or in full if we miss the shipping date we gave you. Buyers in the EU and the UK can cancel for a full refund any time before delivery. To cancel, email support@lucebox.com with your order number. After delivery you can return it within 14 days of delivery, complete and undamaged, with return shipping on you. Outside the EU and the UK a 25% restocking fee is deducted from the refund; buyers in the EU and the UK pay no restocking fee. A unit that arrives damaged, is dead on arrival, or develops a fault covered by the warranty is repaired or replaced at no cost.

Can I run multiple models at once?

Yes. Up to 192 GB of unified memory plus 32 GB of VRAM lets you keep a hot model resident in VRAM while a second sits in unified memory ready to swap in. The CLI handles loading and routing per request.

How does speed hold up with several users?

At 4 to 8 concurrent users, expect each user's generation speed to drop by 10 to 20% at most. If one user gets 100 tok/s on a model, four users each get about 80 to 90 tok/s, roughly 360 tok/s combined. We publish multi-client benchmarks on the blog as they become available.

Your local AI workstation.