Built for inference
128 GB of unified memory holds long context and larger models. 32 GB of GDDR6 keeps the hot path on the GPU.
Our open source inference engine is used by engineers at
Local inference hardware
Built to combine unified memory capacitywith dedicated GPU performance.
Ready in about a minute
Engine, tuned models, and compatible APIs are installed before shipping. Plug in, pair, run.
Private by default
No mandatory cloud path. Your code, prompts, and model weights stay on hardware you own.
$6,499 $7,900
Limited production capacity Tell us what you are building and request a Lucebox.
Apply now Full specs and information
An aluminium chassis engineered for compact size and sustained performance.
Production specification
| Chassis & thermals | |
|---|---|
| Enclosure | Machined aluminium, hand-assembled in-house |
| Dimensions | 340 × 110 × 320 mm (13.4 × 4.3 × 12.6 in) |
| Volume | 11.97 L (approximately 730 in³) |
| Weight | 6 kg (13.2 lb) |
| Cooling | Air-cooled, tuned for low noise under sustained load |
| Connectivity | |
| Ports | 2× USB4, 2× USB-A, HDMI 2.1, 2× DisplayPort 2.1, 5 GbE Ethernet, plus 4 display outputs on the GPU |
| Wireless | Wi-Fi 7 and Bluetooth, built in |
| System | |
| Software | Ubuntu, engine and models pre-installed, OpenAI and Anthropic compatible API |
| Warranty | 1 year, parts and labor, 72-hour burn-in before shipping |
Dense and MoE inference
Performance on Lucebox.
Dense
Qwen3.8-27B
One-user DFlash2 long-prompt decode · 61.3 tok/s on R9700
01
Lucebox / R9700
DFlash2 user run · prose 38.0 · count 217.9
Source / One-user DFlash2 run
- Prefill
- 903 tok/s @ ~6K
- Decode
- 61.3tok/s · long prompt
02
- Prefill
- — not reported
- Decode
- 36.59tok/s
03
- Prefill
- 682.3 tok/s @ 8K
- Decode
- 56.4tok/s
Prompt processing
Point results · prompt lengths differ / tok/s
Token generation
R9700 uses one user's long-prompt result; public workloads differ / tok/s
MoE
DeepSeek V4 Flash
788 tok/s prefill · 86 tok/s decode
01
Lucebox
R9700 + Strix Halo · measured
Source / Lucebox measurement
- Prefill
- 788 tok/s @ 2K
- Decode
- 86tok/s
02
- Prefill
- 1,076.8 tok/s
- Decode
- 35.3tok/s
03
- Prefill
- 790.18 tok/s
- Decode
- 39.35tok/s
Prompt processing
Reported 2K point results / tok/s
Token generation
Published point results / tok/s
Browser-first onboarding
Set it up on lucebox.com. Use it in the apps already on your laptop.
01 / public setup
Open setup Start at lucebox.com/setup.
Pair the nearby machine and put it on your network from the actual onboarding page.
01 / 06 UI and steps are clickable
Private handoff
02 / on your Lucebox
Local / private Finish and connect your apps.
Use the actual local dashboard to prepare the model, enable Lucebox Connect once, and open any supported app.
01 / 10 UI and steps are clickable
- 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 real onboarding UI sends Wi-Fi, account, and optional Tailscale settings directly to the box over encrypted Bluetooth.
- 03
Continue in Manage
The real local dashboard checks the machine, installs its qualified model profile, and starts the private API.
- 04
Connect your apps
Enable the small Lucebox Connect launcher once, then install when needed, configure, and open any supported app from Manage.
Then use Lucebox with the tools you already have
Limited production capacity
Apply now to get yours.
Limited launch price / USD Standard price / USD $6,499 $7,900 $7,90018% launch reductionEnds August 31
Time remaining / --d --h --m --s
01
Share your use case
Tell us what you are building, the workloads involved, how many boxes you need, and your preferred timing.
02
Talk with our team
We schedule a short call to understand the work, confirm technical fit, and discuss delivery needs.
03
Confirm allocation
Capacity is limited and demand is high. If we can allocate a unit, we confirm the delivery window and send a secure purchase link after the call.
1 yearWarranty covering parts and labor
72 hoursFull-system testing before shipping
GlobalShipping included in the price
Tell us what you are building
Required fields marked *Open development
Proudly open source, and built in public.
Public response / X
What builders said after testing our inference engine.
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
01 02 03 04
DFlash
207 tok/s peak / 130 sustained
Speculative decoding on ggml for Qwen3.5-27B Q4_K_M, with the published setup and results.
PFlash
24.8s TTFT at 128K
Speculative prefill cuts a 128K prompt from about 257 seconds to 24.8 seconds on the published CUDA reference run.
Luce Spark
35B MoE / 13.3 GiB
Only the experts used by the current tokens stay resident, so larger MoE models fit on smaller GPUs.
KVFlash
2.9× decode at 256K / 38.6 tok/s
Runs 256K context on a 24 GB card while reducing KV VRAM from 4.6 GiB to 72 MiB through bit-exact paging to host memory.
50+Contributors
500+Pull requests
10+Versions released
Experiments and benchmarks
We share what we learn. Explore every detail in our blog.
01 / PREFILLPFlash: 10× prefill at 128K
About 257 seconds down to 24.8 seconds in the published CUDA reference run, with NIAH retrieval preserved.
Read article →
02 / DECODEDFlash on ggml
Up to 207 tok/s, 3.43× over autoregressive, with 128K context in the published reference configuration.
Read article →
03 / MOELuce Spark on 16 GB
Keep only the experts used by current tokens resident. Self-tuning, one flag, and measured memory use.
Read article →
04 / AMD STRIX HALODFlash + PFlash on Ryzen
Custom decode and prefill paths measured on the Ryzen AI MAX+ 395 against llama.cpp HIP.
Read article →
05 / MOELaguna XS.2 with PFlash
The first MoE target for PFlash, with measured long-context prefill and decode performance.
Read article →
06 / DOCKEROne image for every supported GPU
Prebuilt, self-tuning images across supported NVIDIA GPUs and AMD Strix Halo.
Read article →Questions
Frequently Asked Questions
A plug-and-play computer for local AI inference. It arrives with our open source inference engine, tuned models, custom kernels, and speculative decoding ready to use through a single CLI.
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 KingSpec XG7000 NVMe SSD, and a Corsair SF1000 1000W 80+ Platinum power supply. The custom aluminium chassis has a volume of 11.97 liters.
$6,499 USD per machine through August 31, 2026, then $7,900. Global shipping is included. After a short call confirms fit and availability, we send a secure purchase link.
Production units ship in limited batches. Demand currently exceeds capacity, so delivery windows can vary. We confirm expected timing on a call before sending a purchase link.
Yes. Include the quantity in your application. We review larger requests based on the use case, destination, and current production capacity, so allocation and delivery timing are confirmed on the call.
Yes. Our inference work is public at github.com/Luce-Org/lucebox-hub, with 50+ contributors, 500+ pull requests, and more than 10 released versions. It includes hand-tuned kernels, DFlash speculative decoding, PFlash prefill, and hybrid-model optimizations.
Far more than fits in 32GB of VRAM. The Strix Halo's 128GB of unified memory lets you run much larger models too, like MiniMax 2.7 and DeepSeek V4-Flash. We pre-tune Qwen3.6, GLM-4.6, DeepSeek V4-Flash, and the Llama 4 family, and you can bring your own GGUF, the CLI takes care of loading it.
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.
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.
No. Lucebox runs fully offline once your model is loaded. Internet is only needed for updates and the optional cloud fallback.
The system is quiet at idle and the Radeon AI PRO R9700 cooling system becomes audible during sustained inference. The chassis directs hot air out the back, away from the desk area.
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.
One year on the full machine, parts and labor. We replace or repair any unit that fails under normal use.
Yes. Full root access. It is your machine. Open the shell, install whatever you want, run whatever you want.
If a unit arrives damaged, is dead on arrival, or develops a fault covered by the warranty, contact us and we will arrange the appropriate repair, replacement, or refund. Full terms are provided before purchase.
Yes. Global shipping is included in the listed machine price. VAT, duties, and import taxes remain separate where applicable. We confirm the destination and expected delivery window on the call.
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.