Model · NVIDIA · Text
Nemotron 3.5 Bolt 74B-A9B
NVIDIA's local agentic model, packed into a 24GB download so it runs on a single high-end consumer GPU (32 GB+) — text, image and video input with a 256k context. A compressed version of the 120B-A12B Nemotron 3.5 Super VL, NVIDIA's open data-centre model released the same day.
- Modality
- Text
- License
- OpenMDW-1.1 (Open weights)
- Parameter size
- 74B total / 9B active
- Context window
- 262,144 tokens Up to 256k tokens per the model card; some deployment recipes use shorter windows. Video input up to 2 minutes.
- Released
- October 15, 2026
- Last verified
- October 28, 2026
- Runs locally
- Yes
- Also handles
- vision, reasoning, code
What you need to run it
- Mac
- Yes. Any Apple Silicon Mac (M1 or newer) with at least 32 GB (48 GB recommended) of memory. Newer chips are faster, but memory is what decides whether it runs.
- Windows / Linux PC
- NVIDIA works best. A graphics card with at least 32 GB (48 GB recommended) of video memory.
- No graphics card
- Not practical. Too slow without a graphics card or Apple Silicon.
The 24GB in the name is the download size, not the memory you need: allow 32 GB or more, with NVIDIA recommending a 32 GB+ graphics card; NVIDIA tests it on its own cards and the Mac route is the same llama.cpp file.
Numbers are for: The 2-bit GGUF build of the 74B model (about 24 GB file; the 24GB in the name is the file size). Our estimate from the model’s size — the maker publishes no figure.
Strengths
- Fits in 24GB — native 2-bit quantization-aware training, not a post-hoc quant
- Runs locally through llama.cpp, Ollama, LM Studio and vLLM
- Text, image and video input in one local model
- Leads its class on most agentic benchmarks in NVIDIA's card (Terminal-Bench 2.1 50.98, BrowseComp 66.75, GPQA Diamond 86.43)
- Licence allows commercial use
Weaknesses
- Native 2-bit weights — some fine-tuning and serving toolchains won't handle the mixed IQ2 / NVFP4 format
- Trails Qwen3.6-35B-A3B on PinchBench and on AA-Omniscience accuracy and non-hallucination in NVIDIA's own comparison
- Distilled on about 80B tokens from its parent — expect it to trail Nemotron 3.5 Super VL on knowledge
Try it
| Where | Type | Notes |
|---|---|---|
| Hugging Face | weights | OpenMDW-1.1; GGUF variant available |
| Ollama / llama.cpp / LM Studio | local | 32 GB+ graphics card per NVIDIA (24GB is the download size) |
Official sources
- Model card model
- Announcement announcement
- OpenMDW-1.1 licence docs
Change log
- — Added hardware requirements (Mac support, memory, NVIDIA). Source: Hugging Face card nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF (GPU figures) and file size.
- — Initial entry — first NVIDIA language model in the catalog. Released 2026-10-15 alongside its parent, the open 120B-A12B Nemotron 3.5 Super VL (catalogued here only as context, not as its own entry). 74B total / 9B active MoE in a 24GB native 2-bit pack, 256k context, text/image/video in. Licence OpenMDW-1.1, a permissive model licence (free use, notice-retention, patent-retaliation clause) filed as open-weights because it is not an OSI-approved licence. Hugging Face metadata shows '22B params', which conflicts with the card's own 74B-A9B and likely counts the 2-bit packed tensors; the card's figure is used.
Esc