Back to Home
Pluto Product · Open Source LLMLive on HuggingFace & Ollama

Pluto LM 1.5B

Packaged on HuggingFace & Ollama as Pluto LM

A compact language model fine-tuned for edge inference and IoT agent tasks. Runs on-device. No internet. No data leaves your hardware.

Get the Model

Two ways to run Pluto LM 1.5B locally.

Option 1 · RecommendedOne command via HuggingFace + Ollama

$ ollama pull hf.co/plutoedge/Pluto-LM

Pulls directly from HuggingFace via Ollama. Requires ollama ≥ 0.3.

Option 2Build from Modelfile (PlutoClaw repo)

# Pull base model first

$ ollama pull qwen2.5:1.5b

# Register Pluto LM (domain-tuned for PlutoClaw)

$ cd models/Pluto-LM-v4

$ ollama create plutoedge -f Modelfile

Clone the PlutoClaw repo first, then run from the models directory.

What it does

Edge-Optimized

Quantized for Apple Silicon, Raspberry Pi, and mobile SoCs. Runs in under 2GB RAM.

Privacy-First

All inference runs locally — zero cloud calls, zero data exposure.

PlutoClaw-Native

Designed to plug directly into PlutoClaw agent pipelines as the inference engine.

Open Source

Community-driven development with an open roadmap. Fork it, extend it, ship it.

Performance — what's fast and what's not

Reflex layer (skill & sensor loop)

0.22ms p95measured, Pi 4

Most skills poll sensors every 5–30s and fire actuators on threshold — this loop never calls the LLM. One skill (fire/smoke) also runs an event-driven GPIO path (edge-triggered, gpiozero/lgpio): measured p95 0.22ms / max 9.3ms over N=1000 on a Raspberry Pi 4. Not yet re-verified on Pi 5. Full numbers, caveats, and the reproducible script are in BENCHMARKS.md.

Reasoning layer (Pluto LM, Pi 4 CPU)

~37s

Pluto LM 1.5B reasoning per response on Raspberry Pi 4 CPU, 1024-token context, control-mode prompt. Used for decisions and Q&A — not time-critical sensor loops. Faster on Pi 5 or Mac.

Design principle: time-critical automation never waits on the LLM. The reflex layer acts immediately; Pluto LM judgment runs asynchronously, seconds later, only when needed.

Reproducible Benchmarks

Hardware, script, and measured p50/p95/max latency — one command to reproduce. See BENCHMARKS.md.

Research Paper

The Pluto LM 1.5B architecture, training methodology, and benchmark results will be published as a research paper.

Paper link coming soon