Open Model WeightsNVIDIA · FIELD VERIFIED
Primary sources last checked on 18 Oct 2026. Verification freshness is tracked separately from factual change, and unknown values remain explicit rather than inferred.
FRESHNESS & VERIFICATION
A verification check does not pretend the model changed. Open Model Weights updates the verification date after a real primary-source check, while factual-update dates move only when retained source evidence changes.
CONFLICT INTELLIGENCE
+ Save to my shortlistMonitor changes →Local browser monitoring · no background email alerts
DIRECT ANSWER · FEATURED MODEL BRIEFING
Source-grounded editorial layer · reviewed 2026-10-18
Nemotron 3.5 Bolt IQ2 24GB is a publisher-released open-weight checkpoint in the Nemotron 3.5 Bolt family. This IQ2 model is NVIDIA's native low-bit release targeting a 24GB packed-weight footprint. The claim does not cover KV cache, multimodal input pipelines, runtime buffers or the memory needs of a complete agent. OMW verified its exact publisher repository revision, 91 recognized weight artifacts, openmdw-1.1 license evidence and 262,144 configured tokens on 18 October 2026. This source-based review does not represent an independent model inference, speed or quality benchmark.
WHY IT MATTERS
The native IQ2 release is framed around a tangible 24GB packed-weight target. That is a vendor weight-format goal, not a validated memory budget for a complete production server.
A 74B-total/9B-active hybrid model with image and video processing adds multimodal cache, preprocessing and kernel compatibility requirements.
MODEL-SELECTION VIEW
NVIDIA's official 24GB repository is an IQ2 packed-weight release targeting constrained local operation. A 24GB weight target is not proof that the whole model plus KV cache and media processing fits any particular 24GB consumer GPU.
This IQ2 artifact contains 91 recognized weight entries. Before planning deployment, verify the compatible loader, GPU architecture, visual processor and actual peak memory with your intended context and concurrency.
NVIDIA presents Bolt as a hybrid sparse model with 74B total parameters and about 9B active. Its card describes a Mamba-2/Transformer language backbone alongside a C-RADIO visual encoder for text, image and video inputs.
Active parameters describe the model's per-token execution route, not the storage footprint. Hosting the full model and its multimodal processor is a different problem from running a standalone dense nine-billion-parameter text checkpoint.
The published 24GB variant targets a native low-bit IQ2 representation. Even where packed weights occupy the intended budget, the process still needs cache, activation buffers, visual encoders, temporary tensors and generation workspace.
Do not describe a particular consumer GPU as supported solely because its box advertises 24GB. Reproduce the correct quantized loader on an exact device and driver, then measure realistic peak allocation with the model under load.
OMW's checked configuration records 262,144 tokens for Bolt, with the publisher documenting text, image and video task support. The meaning of a prompt token budget changes with image patches, sampled frames and tool-response history.
For a reproducible agent workload, record image resolution, frame count, document page count and output token budget. Watch for silent truncation and media conversion failures as well as text answer correctness.
The 24GB edition's model-card lineage points to an official BF16 Bolt checkpoint. This link supports a controlled baseline comparison; it does not prove identical task accuracy after quantization or guarantee backend equivalence.
Run the same visual questions, tool-call schemas and text prompts under a compatible engine for each repository SHA. Retain the peak memory and error logs, and keep comparisons explicitly conditional on precision and preprocessing.
NVIDIA declares OpenMDW 1.1 in the source repository metadata. Commercial and non-commercial uses are described as permitted under that agreement's conditions; review the actual terms and any downstream redistribution obligations.
Our editorial judgment: Bolt is particularly relevant when a team can verify the complete deployment against a tight capacity target. Super VL or other larger models may offer different task behavior but bring materially different serving constraints; choose by validated task outcomes.
IMPLEMENTATION & EVALUATION
Reproduction checklist for Nemotron 3.5 Bolt IQ2 24GB, not an OMW inference run. Keep the exact IQ2 checkpoint, tokenizer, media processing, tool harness, software revision and hardware in the report.
Record the official repository SHA, artifact manifest and applicable weight-specific license before benchmarking.
Save a reproducible revision and artifact checksum list together with the exact license source and date.
Model branches may change, and third-party conversions or processors can alter the tested input without notice.
Reproduce the model-card tokenizer, multimodal preprocessing, serving code, precision and tool-call format.
Run at least one positive and one failure-mode task with pinned environment and exact model implementation, or record incompatibility.
Declared library tags do not prove a given package version supports this architecture or compressed format.
Use representative prompt lengths, media dimensions and tool interactions while instrumenting actual memory, latency and correctness.
Retain raw input/output pairs, peak memory, prefill/decode latency, error logs and infrastructure specification.
A configured context limit, weight-only arithmetic or publisher chart is not an independent memory or throughput benchmark.
Run the same acceptance set on a source-linked alternative with aligned prompt budgets and evaluation rules.
Publish both revision IDs, model formats, test harness, pass/fail criteria, observed variance and unsupported cases.
Different engines, sampling strategies, modal input processing and agent permissions confound naive rankings.
Proposed tests, not OMW benchmark results. Read the official model card at the pinned source revision ↗ Inspect OMW field evidence ↗
PRIMARY & INDEPENDENT SOURCES
FAQ
NVIDIA provides a native IQ2 model aimed at a 24GB packed-weight target. OMW has not independently measured peak whole-system memory on a specific 24GB GPU.
The publisher declares 74B total and 9B active parameters. Both figures describe different parts of a sparse model and are not complete GPU memory measurements.
NVIDIA describes text, image and video inputs; the particular processor, backend and frame sampling must be validated in your deployment.
The checked configuration advertises 262,144 tokens. That is a source configuration field rather than an independently measured practical maximum.
NVIDIA declares the OpenMDW License Agreement v1.1. Its commercial permissions remain subject to the full agreement.
The BF16 release provides a higher-precision source reference. Compare task correctness, supported runtimes and complete memory use rather than assuming lower bits always mean better outcomes.
MODEL INTELLIGENCE · SOURCE-FIRST
Open Model Weights records Nemotron 3.5 Bolt IQ2 24GB as a 74B total / 9B active (publisher-declared)-parameter text, image, video → text open-weight model in the Nemotron 3.5 Bolt family, with NVIDIA listed as the developer. The checked source repository provides 91 recognized weight artifacts, format evidence for Safetensors and a verified context window of 262,144 tokens. Values that are not evidenced in the checked sources remain explicitly undisclosed or unknown.
The repository declares NVIDIA OpenMDW License Agreement v1.1. Commercial and non-commercial use declared by NVIDIA; applicable OpenMDW-1.1 model terms still govern. Open Model Weights reports the checked license evidence and classification; this is not legal advice.
The checked sources declare nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 as lineage evidence. Training recipe status: training-described-in-model-card; data disclosure status: dataset-metadata-declared.
DEPLOYMENT PROFILE
These bars visualize weight-only decimal-GB estimates. They exclude KV cache, activations, runtime overhead, optimizer state and sharding.
LICENSE & USE
Commercial and non-commercial use declared by NVIDIA; applicable OpenMDW-1.1 model terms still govern
This section summarizes the checked repository evidence. It is not legal advice.
LINEAGE & PROVENANCE
Open Model Weights does not infer a hidden base model or training corpus when the checked sources do not declare one.
RUNTIME SIGNALS
Declared and model-card-mentioned runtimes are discovery signals. They are not treated as Open Model Weights runtime tests unless explicitly measured.
EVIDENCE TRAIL
This is where the article remains auditable: source links on one side, the registry's observed change history on the other.
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.
01 · WEIGHTS
model-00001-of-00090.safetensorsmodel-00002-of-00090.safetensorsmodel-00003-of-00090.safetensorsmodel-00004-of-00090.safetensorsmodel-00005-of-00090.safetensorsmodel-00006-of-00090.safetensorsmodel-00007-of-00090.safetensorsmodel-00008-of-00090.safetensorsmodel-00009-of-00090.safetensorsmodel-00010-of-00090.safetensorsmodel-00011-of-00090.safetensorsmodel-00012-of-00090.safetensorsmodel-00013-of-00090.safetensorsmodel-00014-of-00090.safetensorsmodel-00015-of-00090.safetensorsmodel-00016-of-00090.safetensorsmodel-00017-of-00090.safetensorsmodel-00018-of-00090.safetensorsmodel-00019-of-00090.safetensorsmodel-00020-of-00090.safetensorsmodel-00021-of-00090.safetensorsmodel-00022-of-00090.safetensorsmodel-00023-of-00090.safetensorsmodel-00024-of-00090.safetensorsmodel-00025-of-00090.safetensorsmodel-00026-of-00090.safetensorsmodel-00027-of-00090.safetensorsmodel-00028-of-00090.safetensorsmodel-00029-of-00090.safetensorsmodel-00030-of-00090.safetensorsmodel-00031-of-00090.safetensorsmodel-00032-of-00090.safetensorsmodel-00033-of-00090.safetensorsmodel-00034-of-00090.safetensorsmodel-00035-of-00090.safetensorsmodel-00036-of-00090.safetensorsmodel-00037-of-00090.safetensorsmodel-00038-of-00090.safetensorsmodel-00039-of-00090.safetensorsmodel-00040-of-00090.safetensorsmodel-00041-of-00090.safetensorsmodel-00042-of-00090.safetensorsmodel-00043-of-00090.safetensorsmodel-00044-of-00090.safetensorsmodel-00045-of-00090.safetensorsmodel-00046-of-00090.safetensorsmodel-00047-of-00090.safetensorsmodel-00048-of-00090.safetensorsmodel-00049-of-00090.safetensorsmodel-00050-of-00090.safetensorsmodel-00051-of-00090.safetensorsmodel-00052-of-00090.safetensorsmodel-00053-of-00090.safetensorsmodel-00054-of-00090.safetensorsmodel-00055-of-00090.safetensorsmodel-00056-of-00090.safetensorsmodel-00057-of-00090.safetensorsmodel-00058-of-00090.safetensorsmodel-00059-of-00090.safetensorsmodel-00060-of-00090.safetensorsmodel-00061-of-00090.safetensorsmodel-00062-of-00090.safetensorsmodel-00063-of-00090.safetensorsmodel-00064-of-00090.safetensorsmodel-00065-of-00090.safetensorsmodel-00066-of-00090.safetensorsmodel-00067-of-00090.safetensorsmodel-00068-of-00090.safetensorsmodel-00069-of-00090.safetensorsmodel-00070-of-00090.safetensorsmodel-00071-of-00090.safetensorsmodel-00072-of-00090.safetensorsmodel-00073-of-00090.safetensorsmodel-00074-of-00090.safetensorsmodel-00075-of-00090.safetensorsmodel-00076-of-00090.safetensorsmodel-00077-of-00090.safetensorsmodel-00078-of-00090.safetensorsmodel-00079-of-00090.safetensorsmodel-00080-of-00090.safetensorsmodel-00081-of-00090.safetensorsmodel-00082-of-00090.safetensorsmodel-00083-of-00090.safetensorsmodel-00084-of-00090.safetensorsmodel-00085-of-00090.safetensorsmodel-00086-of-00090.safetensorsmodel-00087-of-00090.safetensorsmodel-00088-of-00090.safetensorsmodel-00089-of-00090.safetensorsmodel-00090-of-00090.safetensorsmodel.safetensors.index.json02 · LICENSE
Classification describes the checked license terms; it is not legal advice.
03 · HARDWARE
Weight-only decimal-GB estimate from total parameter count; excludes KV cache, activations, runtime overhead, optimizer state and sharding.
04 · FORMATS
“Not observed” means no matching repository artifact/name/dtype signal was found; it does not prove no third-party conversion exists.
05 · LINEAGE
06 · TRAINING ASSETS
07 · RUNTIME SUPPORT
Open Model Weights separates model-repository evidence from external runtime-release evidence. An external support statement does not change the official model repository's weight formats and does not imply an Open Model Weights inference benchmark.
08 · SOURCES
09 · VERIFICATION HISTORY
Checked Hugging Face API metadata, exact repository file list, model card and config when accessible.