Claw Orin packs 67 TOPS of AI compute with 8GB unified memory. It handles 3B–7B parameter models comfortably and is the right choice for personal use, always-on agents, and workloads where privacy and availability matter more than raw parameter count.
Private AI that lives on your desk.
picocluster-claw.com
PicoCluster Claw is a self-contained AI appliance and the reference platform for SMEEP — the Small Model Experimentation and Evaluation Project. Two nodes, one purpose: find out what small models can actually do, on real hardware, in a real agent context.
ClusterClaw runs OpenClaw agents with 28 built-in…
Claw Orin packs 67 TOPS of AI compute with 8GB unified memory. It handles 3B–7B parameter models comfortably and is the right choice for personal use, always-on agents, and workloads where privacy and availability matter more than raw parameter count.
Private AI that lives on your desk.
picocluster-claw.com
PicoCluster Claw is a self-contained AI appliance and the reference platform for SMEEP — the Small Model Experimentation and Evaluation Project. Two nodes, one purpose: find out what small models can actually do, on real hardware, in a real agent context.
ClusterClaw runs OpenClaw agents with 28 built-in MCP tools. ClusterCrush runs local LLMs via Ollama on dedicated NVIDIA CUDA silicon. ThreadWeaver ships pre-installed as your chat interface — multi-provider, branching conversations, full provenance. Plug in, power on, and your private AI stack is running in minutes.
No cloud dependency. No API subscriptions. No data leaving your network. Under $2/month to run, 24/7.
Built for experimentation, not just conversation.
Most private AI hardware is a black box. Claw is the opposite — every benchmark script is open source, every result is reproducible, and the findings are published in plain language.
SMEEP is a living record of what small models can actually do in a real agent context. We run structured benchmarks on the appliance — tool-call ladders, natural-language command suites, quantization comparisons — and publish everything. Current findings include:
* granite4.1:8b is the only model in our dataset to pass every tier of the T1–T4 tool-call ladder (10/10)
* nemotron-3-nano:4b scores 9/10 — the strongest 4B result we've measured
* Q4_K_M format is up to 19× faster than Q4_0 on CPU — same model, same hardware, different quant
* 7B–8B models are a meaningful step up for natural-language agent tasks over the 3B–4B tier
You get the hardware. You get the benchmark suite. Run it yourself, share the results. The more hardware configurations in the dataset, the more useful it becomes for everyone.
Full results and methodology: picocluster-claw.com/reference/benchmarks
Pick your power.
Both configurations ship the same complete stack — OpenClaw, ThreadWeaver, 28 MCP tools, LED feedback, encrypted storage. The difference is inference muscle.
Claw Orin packs 67 TOPS of AI compute with 8GB unified memory. It handles 3B–7B parameter models comfortably and is the right choice for personal use, always-on agents, and workloads where privacy and availability matter more than raw parameter count.
Claw Orin NX steps up to 157 TOPS in Super Mode with 16GB unified memory and an 8-core CPU — 2×-3x the compute and 2× the memory. Run 7B–13B models with room to spare. Built for teams, heavier workloads, and users who want to run the largest models that fit on desktop silicon.
Either way, cloud models stay available when you need them. ThreadWeaver lets you choose local or cloud per conversation. Most users find 70–80% of their work runs perfectly on local models. The rest goes to cloud on your terms, with your API keys, only when you decide.
What's included
* ClusterClaw — Raspberry Pi 5 (8GB), OpenClaw runtime, 28 MCP tools across 5 servers
* ClusterCrush — Jetson Orin Nano Super (8GB) or Orin NX (16GB), dedicated CUDA inference
* ThreadWeaver — pre-installed chat interface, multi-provider, local + cloud
* SMEEP benchmark suite — open-source scripts, ready to run on day one
* Pimoroni Blinkt! LED strip with MCP server for out-of-band status feedback
* Docs, quickstart guide, and community benchmark results at picocluster-claw.com
The Claw stack runs everywhere
OpenClaw and ThreadWeaver aren't locked to the appliance. The same stack installs on macOS (Apple Silicon and Intel), Linux, and Windows via WSL2 — all via Docker Compose, all with identical behavior. One codebase, one git pull to upgrade, four platforms. The PicoCluster Claw appliance is the flagship — dedicated silicon, always-on, pre-configured out of the box — but if you want to run the software on the workstation you already have, you can. Full documentation and install guides at picocluster-claw.com.
