NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB is an open-weight vision-language model packed for a 24 GB label. The model card says it uses native 2-bit (IQ2) quantization to fit in 24 GB and to run alongside other applications. It is a Mamba-2–Transformer hybrid with latent MoE and a C-RADIO v4-H vision encoder, compressed from Nemotron 3.5 Super VL. This site’s calculator does not model that hybrid, so the numbers below are file sizes and the publisher’s deployment notes. They are not a formula fit.
Checked 16 October 2026. Card release line: Hugging Face 15 October 2026. License: OpenMDW-1.1.
Card summary
| Spec | Value |
|---|---|
| Parameters on the card | 74B total, 9B active |
| Precision | IQ2, described as 2-bit with quantization-aware training. Experts are IQ2-XXS at 2.0625 bits |
| Context | Up to 256K tokens |
| Inputs / output | Text, images, and video / text |
| Video | Up to 2 minutes. Card suggests up to 1 FPS / 128 frames at 1080p, and up to 2 FPS / 256 frames at lower resolutions such as 720p |
| Deployment line | 24 GB packed model; vLLM on Ampere (A100), Hopper (H100 80 GB), or Blackwell |
| Parent | Super VL 120B total / 12B active, compressed with Star Elastic. Expert top-K is 9, down from 22 |
The card says Mamba linear layers are NVFP4, and lm_head, embeddings, attention, and latent MoE weights are Q8_0. A LiLiCorr drafter is a separate checkpoint. Its size is not included in the GGUF sum below.
NVIDIA’s blog also says the 24 GB model is on Hugging Face, with support named for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM, and that it runs on GeForce RTX PCs, DGX Spark, DGX Station, and RTX PRO workstations. That is a support list, not a measured memory trace.
Files
GB is 10⁹ bytes. GiB is 2³⁰ bytes. On this site a marketed 24 GB discrete card is 24 GiB. See what fits in 24 GB and the RTX 4090 24GB. DGX Spark is unified memory, not a 24 GB discrete card.
| File or set | Revision | Bytes | GB | GiB |
|---|---|---|---|---|
NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf |
e0fecb69cb3e |
22,018,326,528 | 22.018 | 20.506 |
mmproj-NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB.gguf |
same repo | 1,733,692,736 | 1.734 | 1.615 |
| Those two GGUF files added | 23,752,019,264 | 23.752 | 22.121 | |
| 90 IQ2 safetensors shards | c6cc2e8ffa35 |
24,083,745,496 | 24.084 | 22.430 |
| 62 BF16 weight shards | a6a1e5030819 |
149,774,331,000 | 149.774 | 139.488 |
The mmproj file is the vision projector. It is not inside the language GGUF. The BF16 shard sum does not fit a 24 GiB label. The IQ2 shard sum and the two-file GGUF sum are under 24 GiB and still omit KV cache and runtime overhead. The BF16 repo also has a 136-byte model-summary-idxs.safetensors, which is not in the shard sum.
Safetensors metadata totals: IQ2 repo 22,367,086,600; BF16 repo 74,885,081,096. The first is not the card’s 74B figure.
Bolt E4B
NVIDIA’s 15 October 2026 blog says Nemotron 3.5 Bolt E4B 4GB is coming 23 October 2026, for GeForce RTX, Jetson, and RTX or DGX Spark. Weights and a file table were not published on 16 October 2026. This page does not invent them. The larger open model is Nemotron 3.5 Super VL.
What this page does not claim
No tokens-per-second figure and no benchmark rank are repeated here. Nothing was measured on this site. Qwen3.8-27B remains the kind of open LLM the calculator estimates. Hybrid memory rules are in two GPUs and unified memory. Formula limits are on the methodology page.
FAQ
How much VRAM does Nemotron 3.5 Bolt 24GB need?
NVIDIA's model card says the packed model uses native 2-bit quantization to fit in 24 GB of memory and run alongside other applications. The GGUF language file is 22,018,326,528 bytes (22.018 GB, 20.506 GiB). The separate vision projector is 1,733,692,736 bytes (1.734 GB, 1.615 GiB). Together those two files are 23,752,019,264 bytes (23.752 GB, 22.121 GiB). A marketed 24 GB card on this site is 24 GiB. The file sum is under that label and is still not a runtime estimate. KV cache, activations, and the framework are extra.
Is the Bolt safetensors parameter count 74 billion?
The card says 74B total and 9B active. The IQ2 repo's safetensors metadata total is 22,367,086,600, which is the stored tensor count after quantization, not 74 billion. The BF16 sibling's metadata total is 74,885,081,096. Both are from the Hugging Face API on 16 Oct 2026.
What is the Bolt context length?
The card says up to 256K tokens, shared by the prompt, images, video, tool results, and the reply. Inputs are text, images, and video. Output is text. Video input is up to 2 minutes.
Is Nemotron 3.5 Bolt E4B available?
NVIDIA's 15 Oct 2026 blog says Bolt E4B 4GB is coming 23 Oct 2026. No E4B checkpoint, file size, or parameter table was published on the check date, so this page does not list one.