What I am working on
Most of my AI work this year runs on one machine, an NVIDIA DGX Spark. My Hugging Face profile says it in two words: LLMs and NVIDIA-DGX-Spark. This post is a map of what that one box does. It is still a work in progress.
Quantizing models
The Spark has a GB10 Blackwell chip. I use it to make NVFP4 quants with NVIDIA Model Optimizer 0.41.0. They live in a public collection called NVIDIA-DGX-Spark, for models I re-quantized or altered so they run on one Spark.
- Nemotron-Terminal-32B NVFP4: https://kleinpanic.com/blog/nemotron-nvfp4
- Qwen3-Coder-30B NVFP4, requantized with real C4 calibration, with the router and gate layers left out: https://kleinpanic.com/blog/qwen3-coder-nvfp4
The Nemotron card gives a vllm serve line. NVFP4 kernels need Blackwell, so these quants are for cards like the GB10 and newer.
Training a small model
The Canvas Calendar Agent was trained on the same box: the GB10 with 122 GiB of unified memory, inside an NGC PyTorch container. The DPO step took 9:03. The post is https://kleinpanic.com/blog/canvas-dpo-agent.
Serving retrieval
memory-spark is a memory plugin for OpenClaw agents. Its embeddings, reranker, NER, OCR and speech to text all run as services on the Spark. Its README reports BEIR results with a best NDCG@10 of 0.7889. The post is https://kleinpanic.com/blog/memory-spark.
GSD-OC can use the Spark for semantic retrieval too. Without it, it falls back to BM25 and trigram search. The post is https://kleinpanic.com/blog/gsd-oc.
What is not done
None of this is a product. It is one machine and a lot of experiments. I will write more as the pieces settle.
P.S Thanks for reading. Consider subbing to the RSS feed.
Klein Panic