Descriptions:
Fahd Mirza puts NVIDIA’s newly released Nemotron 3 Embed 1B to a practical cost test, pitting it against Nomic Embed Text in a live agentic retrieval scenario. Running on an RTX 6000 with 48 GB of VRAM and using Qwen 3.8 as the reasoning model, he constructs a 25-chunk corpus of IWC whaling regulations and fires the same three multi-hop questions at both vector stores, logging every search call and token consumed.
The results are concrete: Nomic required 15 searches and burned 12,688 tokens to answer all three questions, while Nemotron 3 Embed finished in 11 searches and just 8,734 tokens — a 31% token reduction from a single swap at the embedding layer. On the hardest question, Nomic spent 4,231 tokens across five searches; Nemotron answered in three searches and roughly 2,110 tokens.
Beyond the numbers, the video contextualizes Nemotron 3 Embed’s benchmark standing: it ranks first at the 1B parameter scale on MTEB with a 72.4% score, supports 34 languages and a 32K context window, and is fully open-weights with commercial licensing. The core argument is that retrieval quality isn’t just an accuracy concern — poor embeddings force agents to loop, burning tokens and inflating costs in ways that compound at scale.
📺 Source: Fahd Mirza · Published August 04, 2026
🏷️ Format: Benchmark Test
