Public Beta

Run AI at the Edge. Privately. Across Any Device.

Inclavate is a distributed inference protocol. Self-host LLMs across your own hardware — phones, laptops, servers — and make speech, music, images and video on it. Nothing leaves your network.

Qwen·DeepSeek·Llama
chat · reasoning · 45 families
Text to video
speech · music · image · video
Mac·Linux·Win
GPU or CPU
OpenAI
compatible API

What can my machines run?

A fishing harbour at dusk with moored boats, lamps along a stone quay and an orange-to-blue sky.

A quiet fishing harbour at dusk, a slow push-in, with a spoken narration.

4 s clip · 1024×576 · LTX-2.3 · M5 Pro 24 GB · 4.3 min

Prompt

A quiet fishing harbour at dusk, seen from the end of a stone quay. Small wooden fishing boats rock gently on calm water, warm yellow lamps along the waterfront glow and reflect in long ripples, the sky fades from soft orange to deep blue, and gulls glide slowly overhead. Slow, steady push-in, cinematic and realistic, gentle natural motion.

Transcript

The tide comes in, and the harbour lights come on.

scroll to explore
Product Preview

A control plane built for
distributed inference operators

Monitor nodes, trace inference paths, manage clusters, and stream live token output — all from one real-time dashboard.

An illustration of the dashboard, with real runs' numbers: Qwen3.5 9B on M5 Pro 24 GB, Wan 2.2 TI2V 5B Turbo on M4 Pro 24 GB.

Open the dashboard →

Text, voice, pictures and video

Made on your machines

51tokens/s

Plan a relaxed three-day trip to Kyoto in autumn. Five short bullet points.

gpt-oss-20b · M5 Pro 24 GB

  • Day 1 – Central Kyoto: Start at Fushimi Inari early, walk to Kiyomizu-dera and Gion for a matcha break; wander the narrow Higashiyama lanes to Yasaka Shrine in the late afternoon.
  • Day 2 – Arashiyama & Sagano: Sunrise stroll through bamboo grove, visit Tenryū-ji temple, lunch by Togetsu-kyo Bridge café, finish with a quiet tea ceremony in a hillside teahouse.
  • Day 3 – Optional Nara or Kyoto finish: Take the early train to Nara for Todaiji & deer park (or stay in Kyoto for Kinkaku-ji and sunset along the Philosopher’s Path).
  • Dining: Choose seasonal kaiseki or izakaya spots that highlight autumn chestnuts, mushrooms, and pickled greens.
  • Travel tips: Stay near central stations to avoid rush trains; use bus passes for scenic routes; schedule breaks at cafés between sights to keep the pace relaxed.

How many times does the digit 7 appear when you write out every whole number from 1 to 1000? Give the number and one sentence of why.

Think · low

5s265 tokens

Answer: The digit 7 appears 300 times when you write out every whole number from 1 to 1000, because in the three‐digit range (000–999) each position—units, tens and hundreds—contains a 7 exactly one-tenth of the time (100 occurrences per position), giving 3 × 100 = 300.

Think · high

112s5,472 tokens

300 – because for every place (hundreds, tens and ones) among the thousand numbers from 000 to 999 each digit appears one-tenth of the time (i.e., 100 times), so there are 3 × 100 = 300 sevens in total.

gpt-oss-20b · M5 Pro 24 GB

A paper boat drifting down a rain-filled street gutter at golden hour, in slow motion.

2 s clip · 1280×704 · Wan 2.2 TI2V 5B Turbo · M5 Pro 24 GB · 7.8 min

Prompt

a paper boat drifting down a rain-filled street gutter at golden hour, ripples and warm reflections, cinematic slow motion

The dripper photo, edited in place: a small potted succulent added beside it.

Z-Image Turbo · M5 Pro 24 GB · 72 s

Prompt

studio product photo of a ceramic pour-over coffee dripper on a glass carafe, on a walnut counter, a small potted succulent beside it, soft morning window light, shallow depth of field

0:00 / 0:11
Transcript

Welcome back. Today we're looking at three small ways to make your morning coffee better, starting with the water: fresh, filtered, and just off the boil.

A short coffee-tips narration in an English voice, Ryan.

Qwen3-TTS · M5 Pro 24 GB · 10 s

Prompt

Welcome back. Today we're looking at three small ways to make your morning coffee better, starting with the water: fresh, filtered, and just off the boil.

0:00 / 0:19
Transcript

City lights are calling me home Neon rivers under my feet We are running, running into the light Hold on, hold on, we own the night

Catchy synth-pop, 118 BPM, female vocal, analog synths and a hooky chorus.

MiniMax-Music3 · M5 Pro 24 GB · 3.2 min

Prompt

Catchy synth-pop, 118 BPM, A major. Bright, polished production: soft analog synth pads, a bouncy synth bass, crisp electronic drums and a sparkling arpeggio. Upbeat, warm, nostalgic, made for a city night. Clear female lead vocal, youthful and open, melodic and articulate, with light reverb that adds space without blurring the words. No heavy vibrato, belting or harsh processing. Start with the arpeggio and a four-on-the-floor beat, bring the vocal in on the verse, lift into a hooky chorus, and end on a clean final beat. Keywords: synth-pop, female vocal, analog synths, bouncy bass, electronic drums, arpeggio, upbeat, catchy hook, nostalgic, 118 BPM, A major.

List every ingredient you can see on this counter, one per line.

Tomatoes Basil Spaghetti Parmesan cheese Garlic clove

Qwen3-VL 8B · M5 Pro 24 GB · 4 s

Across machines

One model, two Macs

Qwen3.6 35B-A3B, a 35-billion-parameter mixture-of-experts model neither Mac can run alone, answering at 9.3 tokens/s — faster than you read.

  • M5 Pro, 24 GB of memory, holds 11.5 GB of the model: layers 0 to 19 of 40, and runs the orchestrator.
  • M4 Pro, 24 GB of memory, holds 9.9 GB of the model: layers 20 to 39 of 40.
  • M5 Pro and M4 Pro are linked over LAN Wi-Fi.
  • Together they run the model at 9.3 tokens/s.

21 GB model24 GB each9.3 tokens/s

A 24 GB Mac can give one model about 19 GB, so each holds half of its layers.

Capabilities

Built for operators who
self-host for privacy

Fully Private

No token ever leaves your network. All inference runs on hardware you control, with encrypted inter-node communication.

Distributed Layers

Split any transformer model across heterogeneous devices — Apple Silicon Macs, GPUs, CPUs — using the Inclavate sharding protocol.

Near-native Speed

Optimized layer routing and LAN-first path selection deliver 30–50 tok/s across a typical 3-node home cluster.

Speech, Music, Images & Video

Narration, cloned voices, songs, pictures and clips, made on the same machines through one job API. Image and video jobs run on whichever node holds the model, several machines at once.

Any Hardware

Run on Apple Silicon, NVIDIA, AMD, or CPU-only nodes. Automatic layer assignment based on available VRAM and RAM.

No Vendor Lock-In

Self-host the entire stack on hardware you own. An OpenAI-compatible API and documented protocol keep your existing tools working — no cloud, no lock-in.

Model coverageView catalog →
157 of the top 200 trending models on Hugging Face
45 architecture families, 124M–1.1T — run models bigger than any single machine can hold.
157 run todayof the top 200
+ 47 media models for images, video, speech and music
38 image & video, 6 speech, 3 music — made on your own machines, through one job API.
  • Llamatext
  • Qwentext
  • DeepSeektext
  • Gemmatext
  • Mistraltext
  • gpt-osstext
  • Phitext
  • GLMtext
  • FLUXimage
  • Stable Diffusionimage
  • Qwen-Imageimage
  • Z-Imageimage
  • Wan 2.2video
  • LTXvideo
  • HunyuanVideovideo
  • Qwen3-TTSspeech
  • MiniMax-Music3music
  • All models in the catalog
Text & understanding
  • Reasoning effort
  • Voice chat
  • Speech to text
  • Image input
  • Audio input
  • Video frames
  • Document OCR
Image
  • Image generation
  • Image editing
  • Inpainting
  • Upscaling
  • OpenAI images API
Video
  • Video generation
  • Image to video
  • Video restyle
  • Longer videos
  • Keyframes
  • Soundtracks
  • Talking portraits
Speech & music
  • Speech generation
  • Voice cloning
  • 11 Indian languages
  • Music generation
Coming nextPDF & text files
Don't see your model? Request it →
Why Inclavate

Own your inference,
end to end

Inclavate
Cloud APIs
Single big GPU
Where your data lives
Never leaves your network
Sent to a third party
Local — one machine
What it makes
Text, speech, music, images & video
Each sold separately
Usually one kind per tool
Model-size ceiling
Bigger than any one box, pooled across nodes
Only what they choose to host
Capped by one GPU's VRAM
What it costs
Hardware you already own
Per token, forever
One expensive GPU upfront
Hardware
Any mix — Macs, GPUs, CPUs
None (their servers)
One high-end GPU
Control
Self-hosted, zero-trust
Their terms & retention
Full, but limited scale
Offline / air-gapped
Yes
No — needs internet
Yes
How it works

From zero to distributed inference
in three steps

01

Deploy a node

Install the Inclavate CLI on any machine. Specify how many transformer layers it should handle based on available memory.

dppan node --layers 0-8 --model llama3.2:1b
02

Join a cluster

Nodes discover each other via mDNS on LAN or libp2p on the internet. The orchestrator coordinates layer chains automatically.

dppan join --cluster local-001 --token <TOKEN>
03

Start inferring

Send prompts via REST or the Chat UI. Your input is sharded through nodes in sequence — no cloud, no data exposure.

curl localhost:8000/v1/chat -d '{"prompt":"..."}'
Quickstart

Running in 60 seconds

bash
$ curl -fsSL https://dl.inclavate.io/install.sh | sh
$ dppan run llama3.2:1b # chat with any model, on this machine
Scale across your devicesFull setup guide →
$ dppan join --orchestrator http://<host>:8000
$ dppan ui --orchestrator http://<host>:8000
OpenAI-Compatible API

Drop-in replacement for any OpenAI client

Inclavate exposes a fully OpenAI-compatible REST API. Swap your base URL — your existing code, libraries, and SDKs work unchanged. Point LangChain, llama-index, or your own app at your local cluster. Speech, music, images and video have a job API on the same server.

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "llama3.2:1b",
    "messages": [{"role": "user", "content": "Hello!"}],
    "stream": true
  }'
Security & Privacy

Your data never leaves
your network

Inclavate is local-first and zero-trust. Run fully self-hosted, or use the managed instance — the Hosted/Local badge in your dashboard always shows exactly where inference is happening.

Private by design

Every token is processed on hardware you control, and inter-node traffic is encrypted. Nothing is sent to a third party — ever.

No telemetry

Inclavate doesn't phone home. No usage tracking, no prompt collection, no analytics baked into the binary.

You hold the keys

Join tokens, PeerID allowlists, and cluster isolation mean only your devices join your network. Zero-trust by default.

ENTERPRISE

Deploy Inclavate inside your org

Private cluster deployments with join token auth, PeerID allowlists, full audit logs, JSONL export, and cluster federation. Your own Inclavate Dashboard — on your infrastructure.

Private cluster isolation
Audit logs (JSONL)
PeerID allowlists
Join token rotation
Air-gap compatible
Read the Docs
Contact

Get in touch

Enterprise deployments, security disclosures, or just questions — we read every message.

Or follow along on X (@inclavateai).

Your AI.
Your hardware.
Your rules.

Start running private, distributed AI — chat, speech, music, images and video — today. Free to run.

Read the Docs