Nemotron 3.5 Bolt
Released on October 15, 2026
Changes from previous version
Same-day smaller sibling of Nemotron 3.5 Super VL, compressed to 74B total and 9B active so a 2-bit checkpoint fits in 24GB. Context is 262,144 tokens. On the BF16 card, Terminal-Bench 2.1 is 53.93 and PinchBench is 90.35.
Release Summary
NVIDIA's open edge model compressed from Nemotron 3.5 Super VL. 74B total parameters, 9B active, text, image, and video in, and a 262,144-token context shared by the prompt and the answer. The 24GB IQ2 checkpoint is built for local agents on GeForce RTX, DGX Spark, and RTX PRO workstations. OpenMDW-1.1 weights also ship in BF16. On the BF16 card, Terminal-Bench 2.1 is 53.93, PinchBench is 90.35, and BrowseComp is 70.00. The 24GB IQ2 checkpoint scores 50.98, 87.82, and 66.75 on those three. No hosted token price is published. Nemotron 3.5 Bolt E4B followed on October 23.
Timeline
Nemotron 3.5 Bolt released
NVIDIA publishes Bolt 74B-A9B weights, including the 24GB IQ2 checkpoint for local agent harnesses.