Running OpenClaw + Ollama on RISC-V: Local LLMs on Open Hardware

The RISC-V Experiment

Jeff Geerling has been pushing RISC-V boundaries — running Ollama, Docker, even GPU drivers on these open-architecture boards. I wanted to see if OpenClaw could run a fully local AI agent on RISC-V. No cloud, no API, just open hardware and open software.

Hardware: SiFive HiFive Premier P550

CPU: SiFive P550 (4 cores, RV64GC)
RAM: 16GB DDR5
Storage: 256GB NVMe
OS: Ubuntu 24.04 (riscv64)
Price: ~$500

Installing OpenClaw on RISC-V

OpenClaw is Node.js — and Node.js runs on RISC-V:

# Node.js has official riscv64 builds since v20
wget https://nodejs.org/dist/v22.22.0/node-v22.22.0-linux-riscv64.tar.xz
tar xf node-v22.22.0-linux-riscv64.tar.xz
export PATH=$PWD/node-v22.22.0-linux-riscv64/bin:$PATH

# Clone and install OpenClaw
git clone https://github.com/openclaw/openclaw.git
cd openclaw && npm install

# It actually works! Native riscv64.
openclaw onboard

Building Ollama for RISC-V

Ollama doesn’t ship RISC-V binaries, so we build from source (thanks Jeff for the guide):

# Install Go
wget https://go.dev/dl/go1.23.linux-riscv64.tar.gz
sudo tar -C /usr/local -xzf go1.23.linux-riscv64.tar.gz
export PATH=$PATH:/usr/local/go/bin

# Build Ollama
git clone https://github.com/ollama/ollama.git
cd ollama
go generate ./...
go build .

# Pull a small model
./ollama pull phi-3:3.8b-mini-4k-instruct-q4_0

Performance Reality Check

Let’s be honest. RISC-V in 2026 is where ARM was 10 years ago for compute workloads:

Model: Phi-3 3.8B (Q4_0)
RAM usage: ~3GB
Tokens/second: 2.1 tok/s
Time to first token: 8.2s

For comparison, a Raspberry Pi 5 gets about 4-5 tok/s with the same model. The P550 is roughly half the speed.

Is It Usable?

For simple tasks? Barely. A response that takes 2 seconds on GPT-5-mini takes 30-60 seconds on local Phi-3 on RISC-V. Tool calling? Hit or miss — Phi-3 at Q4 quantization drops tool call accuracy to about 65%.

# openclaw.yaml — patient mode
models:
  default: ollama/phi-3:3.8b-mini-4k-instruct-q4_0

context:
  maxTokens: 4096  # Keep context small for speed
  compactionThreshold: 3000

Why Do It Anyway?

1. Proving it works — OpenClaw + Ollama on fully open hardware is a statement about software freedom
2. Future investment — RISC-V chips are getting faster every year. What’s 2 tok/s today will be 20 tok/s in 3 years
3. Air-gapped use — no network required, no cloud dependency, no proprietary silicon
4. Education — understanding the full stack from ISA to inference

The Hybrid Approach

In practice, I run OpenClaw on RISC-V with GPT-5-mini (Copilot Pro) as the default, and Ollama as the fallback:

models:
  default: github-copilot/gpt-5-mini
  fallback: ollama/phi-3:3.8b-mini-4k-instruct-q4_0

Best of both worlds: cloud quality when online, local capability when not.

Verdict

Can you run a local AI agent on RISC-V? Yes. Should you? Not yet — unless you’re doing it for the principle. But bookmark this post. In 2-3 years, RISC-V boards with 64GB RAM and vector extensions will make local AI genuinely competitive. And when that day comes, OpenClaw will already run there.