Gemma 4 NVFP4 on the DGX Spark: 271 tok/s at 8 concurrent, native tool calling and reasoning
Native tool calling and reasoning mode on Gemma 4 NVFP4 over 128GB of unified memory.
Hi, I'm Luc Bocahut. I build software, infrastructure, and the occasional agent. This is where I write things down. Some of it is technical, some of it is closer to a manifesto. A few things I have shipped.
Anton, chapter 9: Threads, spawn, and the cast
Concurrency, awakening clones, and a cast of named model specialists working as a team.
Anton, chapter 8: Local LLM optimization, NVFP4 Gemma on DGX Spark
How a billing failure turned into a local-LLM upgrade and a real fallback chain.
A regeneration essay
A longer companion to the manifesto on regenerative software.
A regeneration manifesto
Notes on building software that tries to last.
Anton, chapter 7: Cost, fallbacks, syndic, heartbeat
Two weeks of operational discipline: cost attribution, fallbacks, the syndic domain, and a survey heartbeat.
Anton, chapter 6: The mesh, the sandbox, and self-reflection
Federation, a Deno sandbox, an encrypted secrets table, and a nightly self-review loop.
Anton, chapter 5: LangGraph excised, agents standardized
Replacing a framework with a few hundred lines of runtime, and reshaping the system around two primitives.
Anton, chapter 4: Plumbing matures
A week of turning implicit conventions into explicit contracts: vLLM, LiteLLM, memory, traces, prompts as data.
Anton, chapter 3: Domains widen, browser hardens
Hardening the browser, lifting the work to the LLM, and turning domains into pluggable substrates.
50+ tokens per second on a desktop: running LLMs on the NVIDIA DGX Spark
Standing up a 30B-parameter LLM at 50+ tok/s on the NVIDIA DGX Spark, the technical journey.
Anton, chapter 2: The first weekend
A first full weekend of building turns Anton into something the family can actually use.
Anton, chapter 1: Genesis
Building Anton, a personal agent OS for my family on a DGX Spark, day one.