Blog
Papers, engine internals, and benchmarks on model- and hardware-specific inference.
Quantization
Repairing a Quantized LLM using Weight Space Programming
We gave an LLM agent a poorly calibrated 3-bit checkpoint and NVIDIA's NVFP4 release as a reference. The agent edited 0.0008% of the weights without gradient updates, and LiveCodeBench v6 rose from 30.0% to 47.8%.
Read the post →
Paper
Hibana: Fast Agentic Inference on Unified-Memory Devices
Model surgery, an SSD session cache, and adaptive speculation, co-designed for one DGX Spark. Interactive illustrations of each idea, and results on four models.
Read the post →More posts coming soon.