Model

Nemotron 3.5 Bolt 74B-A9B

OPEN

NVIDIA · Nemotron 3.5 family · released Oct 15, 2026

First model in NVIDIA's edge-focused Bolt line, released October 15, 2026 with Nemotron 3.5 Super VL: a 74B-total / 9B-active hybrid LatentMoE compressed from Super VL and shipped with native 2-bit quantization so it runs in 24GB of memory, plus BF16 reference weights, under OpenMDW-1.1. A smaller Bolt E4B 4GB followed on October 23 (separate entry). Scores are NVIDIA's for the 24GB IQ2 build. NOT FILED: GDPVal-AA-V2 (28.47) is on a different scale from the corpus gdpval-aa column; AA-Omniscience, AA-LCR and WebArena (vision) have no matching column.

ReasoningCodingVisionFunction callingTool useAgentic

1952.2

Elo · rank #99

Parameters

74B

Active params

9B (MoE)

Context

256K tokens

Architecture

Hybrid Mamba-2 / Transformer LatentMoE, 74B total / 9B active, compressed from Nemotron 3.5 Super VL with Star Elastic architecture search and distillation; C-RADIO v4-H vision encoder; 256K-token context; native 2-bit 24GB release

License

OpenMDW License 1.1

Languages

—

API price (in/out)

No hosted API

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

BrowseCompAgents66.8%#31

best: Kimi K3.1 Pro · 92.1%

GPQA DiamondReasoning86.4%#63

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning25.3%#71

best: Kimi K3.1 Pro · 70.0%

MCP AtlasAgents52.2%#11

best: Muse Spark Open · 89.2%

PinchBenchAgents87.8%#4

best: Trinity-Large-Thinking · 91.9%

SciCodeCoding41.0%#6

best: Step 5 Preview · 58.9%

τ³-BankingAgents17.0%#8

best: Step 5 Preview · 42.5%

Terminal-Bench 2.0Coding51.0%#67

best: DeepSeek-V4.1-Flash · 90.6%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

2-bit · GGUF

Fine-tune it

Permissive

QLoRA50.3 GB1× A100 80GB

LoRA157.6 GB1× B200 192GB

Full fine-tune1187.7 GB8× B200 192GB

QLoRA SFT on ~10k samples ≈ $52.04 (1× A100 80GB)

Plan a fine-tune →

Nemotron 3.5 family

Elo progression across releases

Nemotron 3.5 Bolt 74B-A9BOct 20261952.2 Nemotron 3.5 Super VLOct 20262076.9 Nemotron 3.5 Bolt E4BOct 20261569.2

Resources

HFnvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 DOCShttps://developer.nvidia.com/blog/nvidia-nemotron-3-5-super-vl-powers-specialized-agents-with-multimodal-reasoning

API price weights · each benchmark row carries its own source badge (see methodology)