Products

Built on a working lab,
not a slide.

Everything we ship is proven on our own six-system GB10 fleet first — then released as a kit, a tool, a checkpoint, or an application.

Local AI infrastructure

Sparkdeck

DGX SPARK FLEET CONTROL

Unified control for a fleet of DGX Sparks: launch, manage, and stop inference workloads across your systems from one place.

github.com/nero-/Sparkdeck

sparktop

RESOURCE MONITOR

An rtop-style monitor for DGX Sparks — live GPU, memory, and host telemetry for your lab, in the terminal where you live.

github.com/nero-/sparktop

Deployment kits

REPRODUCIBLE MULTI-NODE SERVING

Benchmarked, documented deployment packages that stand open-weight models up on GB10 hardware — single-node and multi-node:

Published checkpoints

JMNI-LABS ON HUGGING FACE

When a hybrid mixture of trained tensors and quantized components outperforms stock releases on GB10-class systems, we publish it ourselves:

JMNI-Labs/Qwen3.8-Flash-Next-NVFP4-QAD5500-Hybrid — a 93B image-text-to-text hybrid for Spark pairs.

Subscription kits — in development

FOR CUSTOMER-OWNED CLUSTERS

Model packs, deployment kits, and rolling updates for clusters you own — designed so a small team can operate a private inference stack without becoming a full-time ops crew. Currently in testing with design partners.

Get in touch for early access →

Recitation

v0.1 — RELEASING SOON

Your personal university professor.

Upload any textbook — PDF, txt, or markdown — and Recitation builds a complete course around it: an auto-generated curriculum, interactive streamed lectures, graded problem sets, and a professor that remembers what you've learned and what you already know.

  • Grounded in your book. Answers cite the pages they came from. Recitation searches the indexed book first, supplements with general knowledge, and never invents page numbers.
  • Voice in and out. Speak your questions (local whisper transcription); hear lectures in human-sounding voices, generated on your machine (Kokoro / Supertonic).
  • Runs entirely on your hardware. Point it at your own cluster, Ollama, LM Studio, or any OpenAI-compatible endpoint. Fully offline-capable once models are on disk.
  • Any model works. No native function-calling required — Recitation supports weak local models as gracefully as frontier ones.

Want it early, or need it for a school or organization? Talk to us.