A 74B total / 9B active hybrid Mamba-2 – Transformer with LatentMoE, compressed out of Nemotron 3.5 Super VL by a method NVIDIA calls Star Elastic rather than trained from scratch. The compression is visible in the routing: expert top-K drops from 22 in the parent to 9. Context is 256K, down from Super VL's 1M.

The name states the target. "Bolt" is the locally-deployable build and "24GB" is the deliverable: with native 2-bit quantization the packed model fits a single consumer GPU, which is the whole point of taking a 120B model apart. That makes it the interesting half of the pair — Super VL is the capability statement, Bolt is the claim that the capability survives being squeezed onto hardware people already own.

Published scores: GPQA Diamond 86.43 (against the parent's 89.52, so roughly three points of reasoning for a 38% parameter cut and 2-bit weights), PinchBench 87.82, Terminal-Bench 2.1 50.98, WebArena vision 50.86, SciCode 41.02, GDPVal-AA-V2 28.47. Training cutoff May 2026, OpenMDW-1.1, with GGUF and a LiLiCorr variant alongside.

Artificial Analysis scored it on release day at 21 on Intelligence Index v4.3.2 (21.4 unrounded), #4 of 67 open-weights models in the 40B–150B class and first among US models in it. The compression claim holds up on AA's own axis: it lands 1.5 points behind Nemotron 3 Ultra on about a seventh of Ultra's parameters. The cost shows up elsewhere — AA measures it as very verbose, 230M output tokens against a 100M median for comparable models, and 82k per index task against Super VL's 46k. So the intelligence survives the squeeze better than the token budget does.

Ten days earlier, on October 5, NVIDIA had put out a much smaller sibling under the same name. Nemotron-3.5-Bolt-E4B is not a Nemotron pretrain: the card says it is "further post-trained from Gemma-4-E4B-it," at 8B total / 4B effective parameters with 64K input and output. It is tuned for multi-turn tool use, instruction following and grounding in local voice assistants and on-device personas. It ships as a 4GB Q4_K_M GGUF for llama.cpp, a BF16 checkpoint and DFlash speculative drafters, under OpenMDW-1.1. On NVIDIA's own table the BF16 build beats Gemma-4-E4B-it on MultiChallenge (53.10 vs 42.08), IFBench (63.79 vs 42.84) and Tau2-Bench (55.10 vs 38.50), but trails Qwen3.5-4B on Tau2-Bench (75.06) and RULER 64K. Artificial Analysis lists it as nemotron-3-5-bolt-e4b-4gb with no published Intelligence Index score yet.

Model Details

Architecture MOE
Parameters 74B
Active params 9B
Context window 262,144
AA Intelligence 21 on v4.3.2
License OpenMDW-1.1
Base model nemotron-3.5-super-vl

Benchmark Scores

Benchmark Score Mode
PinchBench 87.82 —
GPQA Diamond 86.43 —
Terminal-Bench 2.1 50.98 —
WebArena (vision) 50.86 —
SciCode 41.02 —
GDPVal-AA-V2 28.47 —

Variants

Name Parameters Notes
Bolt 74B-A9B 24GB — native 2-bit, packs into 24GB for a consumer GPU
Bolt 74B-A9B BF16 — full-precision reference weights
Bolt 74B-A9B 24GB GGUF — GGUF build
Bolt E4B 4GB / E4B BF16 — 2026-10-05; post-trained from Gemma-4-E4B-it (not from Super VL), 8B total / 4B effective, 64K in/out; 4GB = Q4_K_M GGUF; DFlash drafters alongside; AA slug nemotron-3-5-bolt-e4b-4gb unscored
open-weightsmoeefficiencyedgequantizationreasoning

Related