Your data stays private
Client files, code and unreleased ideas never reach a cloud model. Runs fully offline, stores no prompts, needs no account with us.
Partners
Hardware
Private, always on, ready out of the box.
- Unified memory
- 128 GB
- AMD Ryzen AI MAX+ 395,
LPDDR5X-8000 at 256 GB/s - Dedicated GPU
- 32 GB
- AMD Radeon AI PRO R9700,
GDDR6 at 640 GB/s - Storage
- 2 TB
- NVMe, up to 7,400 MB/s read
- Connectivity
- 5 GbE
- Wi-Fi 7, Bluetooth and 2× USB4
- Power supply
- 1,000 W
- 80 PLUS Platinum, SFX
- Warranty
- 1 year
- Parts and labor,
after 72 hours of burn-in
Built for coding agents
Point Claude Code, Codex, OpenCode or Pi at its Anthropic and OpenAI compatible API. Repos, tickets and docs stay on your network.
Always on, no meter
Research bots, automations and agent teams, running 24/7. Pay once, share it with your team, and forget per-token bills and rate limits.
Optimized for speed from the start
The engine, custom kernels, speculative decoding and tuned models come installed for this exact hardware. Plug it in, pair it, and it runs at full speed.
Performance
Two processors, one engine, built for speed.
Over 2× the reported DGX Spark decode speed.
Dense
Qwen3.8-27B
- Prefill
- 903 tok/s
- Decode
- 61.3 tok/s
From one user's DFlash2 run on a long prompt. DGX Spark and Mac M5 Max are public records with their own setups.
DetailsMoE
DeepSeek V4 Flash
- Prefill
- 788 tok/s
- Decode
- 86 tok/s
Measured on Lucebox at 2K context. DGX Spark and Mac M5 Max are public records with their own setups.
Details
Setup
Easy setup. Use it in the apps on your laptop.
01 / 11 UI, tabs and steps are clickable
Setup
Manage
- 01
Open the setup page
Visit lucebox.com/setup from a nearby laptop or Android device. There is no app or CLI to install.
- 02
Pair and connect
The onboarding page sends Wi-Fi, account, and optional Tailscale settings to the box over encrypted Bluetooth.
- 03
Continue in Manage
The local dashboard checks the machine, installs the qualified model profile, and starts the private API.
- 04
Connect your apps
Enable Lucebox Connect once, then open any supported app from Manage with a single click.
- 05
Monitor your Lucebox
Tokens per app and model, recent requests, and live machine meters. Prompts and replies are never stored.
Works with the tools you already use
Order
Order your Lucebox.
Developer
For individuals$5,999€5,999 USD per machine / shipping includedEUR + VAT per machine / shipping included
- Lucebox Engine, open source, with tuned models preinstalled
- 1-year warranty on parts and labor
- Worldwide shipping included
Business
For teams and companies12 months of Lucebox Engine Pro included
$7,999€7,999 USD per machine / shipping includedEUR + VAT per machine / shipping included
- Everything in Developer
- White-glove install and onboarding
- Direct chat support with the engineering team
Community
Proudly open source, and built in public.
Ahmad@TheAhmadOsman
I like your stuff so far, keep going
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.
Lotto@LottoLabs
Interesting run w/ Dflash from the lucebox-hub guys
Rijndael@rot13maxi
speculative PREFILL?????
Riku Pasonen@Raitziger
I have tested some LLM server software for home PCs for Linux and Windows. Fastest and best for running home is Linux running 145 t/s, Lucebox. @pupposandro @luceboxai
fahd Mirza@fahdmirza
PFlash just killed the 4-minute blank screen problem. 128K token prefill in 25 seconds, same GPU, same model, no compromises
Geek Lite@QingQ77
Consumer-grade GPUs actually have sufficient hardware potential, general-purpose frameworks just waste most of it on overhead. Lucebox releases that potential through hand-written kernels, letting even a 2020 RTX 3090 rival Apple's latest chips on efficiency.
Takyon∞@Takyon
Crazy I was litteraly wondering how can I increase my token speed 10 min ago
Kyz2ren@ky2renzz
Crazy what @pupposandro just dropped on Qwen3.5-27B. 207 tok/s on a single 3090 with Q4_K_M and full DFlash speculative? Chinese labs + ggml hacks are just cooking on consumer hardware right now. This is the kind of local win I like to see.
Ivan Fioravanti@ivanfioravanti
RTX 3090 ready for a new life! Bringing it to @luceboxai team to make some experiments together
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
CAPET@Capetlevrai
This is very very good work BRAVO. I love it
nash_su@nash_su
Nearly 10x faster! After finishing Decoding, it starts cranking through Prefill. The previous DFlash was already stunning enough, and now they've added PFlash. Speculative prefill, up to 10x speedup. Go try it right now.
AJ@ItsmeAjayKV
First time reading about speculative prefill, and it's crazy. 257s down to 24s for a 128K prompt on a single RTX 3090. Great article, definitely go ahead and give this a read.
Twon.@Web3Twon
impressive... so this is what it looks like when you focus on a set ram limit and optimizing for a single model above everything
Our open source inference engine is used by engineers at
55+Contributors
685+Pull requests
80K+Model downloads
Blog
We share what we learn. Explore every detail in our blog.
AMD · VISION
Vision LLM inference: Lucebox has 3.2x the throughput of NVIDIA DGX Spark
At the highest load each machine served, one Lucebox answered 58 image questions a minute (Qwen3.8-27B on its AMD Radeon AI PRO R9700 and the 284B DeepSeek V4 Flash Vision on its Strix Halo, at the same time) and an NVIDIA DGX Spark running llama.cpp 18. With the same Qwen model file and eight questions at once, the R9700 alone finishes 2.4x sooner.
AMD · COOLING
Inference-aware cooling: 29% less fan speed on the AMD R9700, same temperatures
AI inference and cooling designed together on the R9700: the model forecasts the work ahead, a learned thermal model picks the slowest safe fan speed. Chat silent, agent loops 29% and batch work 25% quieter, same peak temperatures and throughput.
AMD · STRIX HALO
ROCm beats Vulkan on Strix Halo
DeepSeek V4 Flash on one Strix Halo, both engines on the same box on the same day: Lucebox ROCm is 17 to 47% faster on prefill and 28 to 47% faster on speculative decode than llama.cpp Vulkan at 8K, 32K and 123K, with plain decode measured on the same box too.
AMD · DEEPSEEK
3.63× the DGX Spark on DeepSeek V4 Flash
Asymmetric parallelism splits the full 284B model across the R9700 and the Strix Halo: 3.63× the decode speed of one DGX Spark, for 36% less than two.
LONG CONTEXT
KVFlash: 256K context with 72 MiB of KV
Cold 64-token KV chunks page to host RAM bit-exact, so decode holds at 38.6 tok/s from 64K to 256K with unchanged accuracy.
AMD · MOE
DeepSeek V4 Flash on the Ryzen AI MAX+ 395
The full 284B model runs from 128 GB of unified memory: up to 32 tok/s decode and roughly 250 tok/s indexed sparse prefill.
FAQ
Frequently Asked Questions
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
$5,999 USD per machine for the Developer tier (individuals) and $7,999 USD for the Business tier (teams), which includes one year of Lucebox Engine Pro. Worldwide shipping is included. Import duties and taxes are not: whatever your country charges is billed to you by the carrier, usually before delivery. You pay at checkout on shop.lucebox.com.€5,999 plus VAT per machine for the Developer tier (individuals) and €7,999 plus VAT for the Business tier (teams), which includes one year of Lucebox Engine Pro. Shipping is included and VAT is added at checkout. Orders ship from Italy, so EU deliveries pay no import duties or customs fees. You pay at checkout on shop.lucebox.com.
On shop.lucebox.com, by card or with 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 shipping. EU orders add VAT at checkout; deliveries outside the EU may owe import duties and taxes, billed by the carrier.
An AMD Ryzen AI MAX+ 395 with 128GB of LPDDR5X-8000 unified memory, an AMD Radeon AI PRO R9700 with 32GB of GDDR6, a 2TB NVMe 1.4 SSD, and a 1000W 80+ Platinum SFX power supply. The extruded aluminium chassis, anodized black, has a volume of 11.97 litres.
Orders placed now ship in November 2026, due to high demand. The shop emails your receipt right away and we email tracking the day your machine leaves.
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
The service tier around the open source engine: white-glove install, direct chat support with our engineers, priority delivery, custom model tuning, implementation help, and security hardening. It costs $199 per month per machine; one year is included with the Business tier and it can be added to any Lucebox.The service tier around the open source engine: white-glove install, direct chat support with our engineers, priority delivery, custom model tuning, implementation help, and security hardening. It costs €199 per month per machine; one year is included with the Business tier and it can be added to any Lucebox.
Far more than fits in 32GB of VRAM: the engine runs one model across the R9700 and the 128GB 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
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.
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.
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.
Yes. Full root access. It is your machine. Open the shell, install whatever you want, run whatever you want.
Yes. We ship worldwide from Italy, and shipping is included in the price. Inside the EU there are no import duties or customs fees, and VAT is added at checkout. Outside the EU, import duties, taxes and customs fees are not included: whatever your country charges is billed to you by the carrier, usually before delivery. The shop collects your shipping address at checkout.
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.
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 performance page collects the published comparisons.
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.
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.
No. Lucebox runs fully offline once your model is loaded. Internet is only needed for updates and the optional cloud fallback.
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.
1 year on the full machine, parts and labor. We replace or repair any unit that fails under normal use.
You can cancel for a full refund at any point before your machine ships: 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.
Yes. The 128 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.
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.