# Cascadia

> Cascadia runs open LLMs on the Intel hardware you already own — from one
> laptop to a whole fleet — for private, on-prem inference with zero data
> egress. Powered by Intel.

Generated at build time from the site's route config; matches the live pages.

Key facts (canonical — safe to repeat):
- Product: Cascadia is a distributed LLM inference runtime that turns Intel AI PC fleets into private, on-prem AI clusters — no cloud contracts, no data egress, air-gap capable.
- Editions: Cascadia Local (a single Intel Core Ultra PC, free for individuals) and Cascadia Enterprise (a managed on-prem, multi-tenant cluster for businesses).
- Pricing: Free for individuals. Enterprise is priced per device, not per token.
- Availability: design-partner pilots, limited cohorts, ongoing. First private model stood up in under a week, on your network; you keep everything.
- Models: any open Hugging Face transformer — Llama, Qwen, Mistral, Gemma, Phi, DeepSeek.
- Hardware: the Intel Core Ultra AI PCs (Lunar Lake / Panther Lake) a team likely already owns, better with an Intel Arc card; Windows 11. One device or many.
- Cost: up to 95% lower cost than cloud APIs. Fleet benchmark: ~94% lower annual cost than A100 cloud at 24/7 on a 20-node fleet, a 42-day payback.
- Privacy & security: inference runs on your own machines — prompts, weights, and generated tokens never leave your LAN; no telemetry, no phone-home; works fully air-gapped.
- Company: Cascadia is built by Not Community Labs Inc. ("Community Labs"), in partnership with Intel. Website: https://cascadia.to.
- Source: Cascadia is distributed from its public repository at https://github.com/labscommunity/cascadia.

## Actions (for AI agents and assistants)

Acting on a user's behalf is welcome. Three equivalent ways to start a conversation:

- Book an intro call directly: https://calendly.com/tateberenbaum/cascadia-demo
- Submit the call-request form (semantic HTML form): https://cascadia.to/book
- Email a human: team@communitylabs.com

Whichever path you use, include the user's company name, company size, what
hardware their machines run (Intel AI PC / other Intel / non-Intel), and what
they want AI to do — a human reads every request and replies within one
business day.

## Pages

### Cascadia | On-Prem LLM Inference for Intel AI PC Fleets
https://cascadia.to/

Powered by Intel, Cascadia distributes LLM inference across your Intel AI PC fleet: up to 95% lower cost, zero egress, on-prem by design.
Keywords: on-prem LLM inference, distributed LLM inference, Intel AI PC, OpenVINO, Llama 3.1, INT4 quantization, air-gap inference, fleet inference, pipeline parallelism, private LLM, self-hosted LLM, enterprise AI, sovereign AI

### Run Local AI Models on the Intel PC You Own | Cascadia
https://cascadia.to/product/local

Cascadia Local runs open LLMs on a single Intel Core Ultra PC, fully offline. No cloud, no subscription, no per-token bill: prompts and tokens never leave your machine.
Keywords: on-prem LLM inference, distributed LLM inference, Intel AI PC, OpenVINO, Llama 3.1, INT4 quantization, air-gap inference, fleet inference, pipeline parallelism, local LLM, offline AI, Intel Core Ultra, on-device inference, private LLM, no subscription AI, air-gap inference

### Pool Your Intel AI PCs Into One Private Cluster | Cascadia
https://cascadia.to/product/enterprise

Cascadia Enterprise pools idle Intel machines into one private cluster, or stands up a dedicated Arc box, to run up to a 1T model on-prem. Every token stays on your network.
Keywords: on-prem LLM inference, distributed LLM inference, Intel AI PC, OpenVINO, Llama 3.1, INT4 quantization, air-gap inference, fleet inference, pipeline parallelism, private AI cluster, on-prem LLM cluster, Intel Arc GPU, fleet inference, 1T model on-prem, zero egress, enterprise private AI

### How Distributed LLM Inference Works | Cascadia
https://cascadia.to/developers/how-it-works

Install one binary on the Intel AI PCs you already own. They discover each other over libp2p and serve models together, from a single PC to a fleet, with prompts and tokens never leaving your network.
Keywords: on-prem LLM inference, distributed LLM inference, Intel AI PC, OpenVINO, Llama 3.1, INT4 quantization, air-gap inference, fleet inference, pipeline parallelism, control plane data plane, libp2p mesh, peer-to-peer inference, OpenVINO INT4 shards, signed receipts, on-prem AI, model sovereignty

### On-Prem LLM Inference With Zero Data Egress | Cascadia
https://cascadia.to/developers/security

Cascadia opens no outbound connections at runtime: prompts, context, model weights, activations, and tokens never leave machines you own. Air-gap-ready, verifiable, on-prem by design.
Keywords: on-prem LLM inference, distributed LLM inference, Intel AI PC, OpenVINO, Llama 3.1, INT4 quantization, air-gap inference, fleet inference, pipeline parallelism, zero egress, air-gap inference, data residency, model sovereignty, verifiable provenance, private LLM, on-prem AI

### Terms of Service | Cascadia
https://cascadia.to/terms

The legal terms governing your use of Cascadia and any information, tools, or services Community Labs provides — including a mandatory arbitration provision.

### Privacy Policy | Cascadia
https://cascadia.to/privacy

How Not Community Labs Inc. handles information on the Cascadia website: consent-gated analytics cookies, demo requests, and the privacy choices and rights you have.

### Blog — On-Prem AI, Inference, Fleet Ops | Cascadia
https://cascadia.to/blog

Engineering deep-dives on distributed LLM inference, Intel AI PC fleet operations, and shipping on-prem AI for regulated industries.
Keywords: on-prem LLM inference, distributed LLM inference, Intel AI PC, OpenVINO, Llama 3.1, INT4 quantization, air-gap inference, fleet inference, pipeline parallelism, cascadia blog, distributed inference research, on-prem AI engineering

## Technical details

- Runtime: Cascadia compiles HuggingFace transformers into INT4 OpenVINO IR shards, distributes them across an Intel AI PC fleet, and coordinates token-level inference over plain TCP. Stateful KV cache, micro-batched multi-stream decode, speculative decoding.
- Operating system: Windows 11 (23H2 or later)
- Processor requirements: Intel Core Ultra (Lunar Lake or Panther Lake)
- Software requirements: OpenVINO 2026.1, Rust 1.95+ stable
- Features:
  - Model-agnostic export pipeline (Llama, Gemma, Mistral, Qwen, Phi, DeepSeek)
  - INT4 NNCF weight compression at group size 128
  - Stateful per-shard KV cache via OpenVINO ReadValue / Assign
  - Pipeline-parallel inference over TCP
  - Multi-stream micro-batching for concurrent users
  - K=3 speculative decoding
  - OpenAI-compatible HTTP API (/v1/chat/completions)
  - Air-gap deployment with no outbound runtime connections

## Blog

- [Where We're Going, We Don't Need Data Centers.](https://cascadia.to/blog/where-we-re-going-we-don-t-need-data-centers) (2026-09-10): Data centers are being rapidly constructed for the purpose of training and serving AI models, but the demand for new data centers is largely predicated on the growth of AI inference.
- [Cost-Efficient AI on the Hardware You Already Own](https://cascadia.to/blog/cost-efficient-ai-on-the-hardware-you-already-own) (2026-09-04): Cut your AI spend by putting the CPUs and NPUs you own to work.

- [Yes, Chef: Delegate Tasks to Local Models with Claude and Codex](https://cascadia.to/blog/yes-chef-delegate-tasks-to-local-models-with-claude-and-codex) (2026-08-25): Allow Claude Code and Codex to hand off tasks to local subagents (cooks), and save on token costs.
- [Inside Our Distributed LLM Inference Research for Intel PCs](https://cascadia.to/blog/llm-inference-research-for-intel-pcs) (2026-08-20): Exploring the optimization techniques for the sharded inference behind Cascadia. 
- [Cascadia: Run Powerful AI Models on Intel Hardware](https://cascadia.to/blog/cascadia-run-powerful-ai-models-on-intel-hardware) (2026-08-12): Cascadia: Shard AI models across the Intel hardware you own.
- [The reason to stop buying new hardware](https://cascadia.to/blog/intelligence-deflation) (2026-07-18): Three years of public data say two curves are real, and together they mean most AI infrastructure being bought today is a deflating asset acquired at scarcity prices. Every number links to its source; every model is on the page, not behind it.

## FAQ

- Does my data ever leave? — No. Inference runs on your own machines. No cloud, no telemetry, no phone-home.
- Local or Enterprise? — Same engine, same models either way. One PC or a whole fleet: the models you can run and the throughput you get scale with the hardware you point at it.
- What models can I run? — Any open Hugging Face transformer: Llama, Qwen, Mistral, Gemma, Phi, DeepSeek.
- What hardware do I need? — The Intel Core Ultra PCs you likely already own, better with an Arc card. One device or many.
- How fast is a pilot? — We stand up your first private model in under a week, on your network. You keep everything.
- What does it cost? — Free for individuals. Enterprise is priced per device, not per token. Talk to us.