Nemotron 3.5 Bolt 74B — VRAM, speed & local setup
Written by Jakub Rusinowski · Last updated
NVIDIA's local-agent model for RTX Spark laptops, compressed from Nemotron 3.5 Super VL (120B-A12B): 74B parameters, 9B active (top-9 experts), in a hybrid stack of 40 Mamba-2, 40 latent-MoE and only 8 attention blocks, so its KV cache is about a tenth of a same-depth transformer's. Takes text, images and video; 262,144-token context. The official local build is a 2-bit GGUF (experts at IQ2_XXS, recovered with quantisation-aware distillation): 22.0 GB, plus a 1.7 GB vision projector. NVIDIA calls it the "24GB" version, but its own card lists RTX GPUs with 32 GB or more, with 28 GB and 32 GB recipes; on a 24 GB card it is tight. Ships with a LiLiCorr speculative-decoding drafter.
Nemotron 3.5 Bolt 74B needs about 23 GB of VRAM at IQ2_XXS QAD — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
VRAM and speed by quantization
Quoted against NVIDIA RTX 4090 (24 GB). Includes the KV cache at 8K context, so it reads higher than the headline figure.
| Quant | Memory | VRAM | Speed (est.) | Fit |
|---|---|---|---|---|
| IQ2_XXS QAD 2.38 bpw | 22.9 GB | ~171 tok/s | Tight | |
| BF16 16.00 bpw | 150.6 GB | — | Too big |
Black marker = usable memory on the NVIDIA RTX 4090 (24 GB). Estimates from the memory-bandwidth roofline on the methodology page. Nemotron 3.5 Bolt 74B VRAM calculator →
Get Nemotron 3.5 Bolt 74B running
The cheapest catalogued GPU that runs Nemotron 3.5 Bolt 74B is the NVIDIA GeForce RTX 5090 (32 GB).
or compare on Vast.ai from $0.35/hr (typical low · varies)
As an Amazon Associate we earn from qualifying purchases. Cloud GPU links are referral links — we may earn a commission at no extra cost to you.
How to run Nemotron 3.5 Bolt 74B
Install Ollama, then run:
ollama run nemotron-3.5-bolt:74b-a9bSpecifications
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source. Still unconfirmed: arch.kv_layers.
- Parameters
- 74 Billion (9B active)
- Context window
- 262,144
- Architecture
- Hybrid Mamba-2 + latent MoE + attention (8 of 88 blocks attend), vision encoder
- Provider
- NVIDIA
- Licence
- OpenMDW-1.1
- Availability
- Weights downloadable
- Practical deployment
- Single consumer GPU
- Specified at
- IQ2_XXS QAD
- Vision projector
- +1.7 GB when you send images (not included above)
- System RAM
- 32 GB
- Record updated
- 2026-10-15
Commercial use permitted. No usage restrictions beyond attribution.
Limits and caveats
- The "24GB" name is the file footprint (22.0 GB weights + 1.7 GB vision projector). NVIDIA's own deployment table starts at 32 GB GPUs.
- Only one local quantisation is published (NVIDIA's 2-bit QAD build); the BF16 checkpoint is meant for fine-tuning, not local inference.
- LM Studio support is reported in launch coverage but not on the model card.
See also Nemotron 3.5 Lightning 30B-A3B
Quality and use cases
Scores as published by the model’s authors or an independent evaluator — quality, not throughput, and not measured by us.
Can I run Nemotron 3.5 Bolt 74B on my GPU?
Other Nemotron 3.5 sizes
Nemotron 3.5 Bolt 74B — frequently asked questions
How much VRAM does Nemotron 3.5 Bolt 74B need?
About 23 GB at IQ2_XXS QAD — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Nemotron 3.5 Bolt 74B run on an RTX 4090 (24 GB)?
Yes. Nemotron 3.5 Bolt 74B needs about 23 GB at IQ2_XXS QAD, inside a 24 GB card, at an estimated 171 tokens/sec.
How do I run Nemotron 3.5 Bolt 74B locally?
Install Ollama and run `ollama run nemotron-3.5-bolt:74b-a9b`. That pulls the weights and starts a local OpenAI-compatible endpoint; after the download nothing leaves the machine.
What other sizes does Nemotron 3.5 come in?
Nemotron 3.5 Lightning 30B-A3B (19 GB), Nemotron 3.5 Bolt 74B (23 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.