Open research artifacts, four-axis carded.

Every artifact here ships with the same disclosure shape — quality, throughput, vertical skill, thermal envelope — and was measured on a single DGX Spark. Browse by kind.

Quantizations 06

Quantized base models, Spark-measured across perplexity, throughput, vertical eval, and thermal envelope.

Browse quantizations
LoRA adapters 01

LoRA fine-tunes with training receipts, evaluation deltas, and bounded drift disclosures.

Browse lora adapters
Adapters Coming soon

Prefix, prompt, and IA³ adapters for specialized inference.

Datasets Coming soon

Curated and synthesized training/eval datasets.

Benchmarks 02

Open benchmarks and eval harnesses with publishable scores.

Browse benchmarks
Notebooks 05

Builder + user Jupyter notebooks, one-click Colab/Kaggle on-ramps for every vertical release.

Browse notebooks
Harnesses 04

Reproducible Spark-agent harness profiles — serving lane, throughput, tool-call reliability.

Browse harnesses
Skills 01

Agentskills.io-format skill bundles — portable across Claude Code, Hermes, and the rest.

Browse skills
Apps 01

Operator apps that run on your own DGX Spark — local-first cockpits over the models, datasets, benches, and harnesses you've produced.

Browse apps

Every artifact is published alongside a field note. Read the running log of releases at /field-notes/.