NVIDIA took OpenAI's model and distilled it—cutting it down from 120B to 88B, with inference speed boosted by 2.82x. This model is called gpt-oss-puzzle-88B and just dropped on Hugging Face. Let me explain why this is more important than it seems. The technology used is called Puzzle, a post-training neural architecture search framework. Simply put—after the model is trained, NAS is used to trim redundant structures. The result? 88B parameters running on a single H100 achieved 2.82x the throughput of the 120B version. Accuracy didn’t drop; in some scenarios, it even improved—maintaining 108.2% accuracy under low inference budgets. What does this mean? Same hardware, faster model, same or even better results. GPU costs slashed directly. What does this mean for the market? NVIDIA is doing something very smart. It’s not just selling chips—it’s showing you that with its chips + optimization tools, your inference costs can plummet. This is an upgraded version of the "selling shovels" logic: "selling shovels + teaching you how to dig faster." For developers, the 88B model is open-source and ready to use, supporting vLLM and Transformers with 128K context. No need to wait for APIs—you can deploy it yourself. My take—model "slimming" will be the main theme by 2026. It’s not about whose model is the biggest; it’s about who can achieve the best results with the least compute power. NVIDIA has already picked a side: inference efficiency > parameter size. If this direction works out, everyone running local models will benefit.