NVIDIA-Nemotron-3.5-Lightning-30B-A3B GGUF size and VRAM requirements
ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF is a very large language model with 31.58 billion parameters, built on the nemotron_h_moe architecture. It is released under the other license and has been downloaded 107,414 times.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF is a Mixture-of-Experts model with 128 experts, of which 6 are active on each token. Routing only 6 of 128 experts makes it noticeably faster than a dense model of the same size — but every expert still has to be held in memory, so the numbers below are set by the full parameter count, not the active one. It is built from 52 transformer layers, a hidden size of 2,688, 32 attention heads. It uses a hybrid attention design: only 6 of its 52 layers keep a full KV cache, while the rest use linear attention with a constant-size state. Long contexts therefore cost far less memory than on a conventional model of this size.
To run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF locally at a 4,096-token context, its quantized versions need between 18.42 GB (Q4_0, lowest quality) and 59.67 GB (BF16, highest quality) of memory, weights plus KV cache and a system margin included. Context length drives that memory directly: at 4,096 tokens the KV cache for NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF is about 0.02 GB, rising to roughly 6.0 GB at its full 1,048,576-token context. Shorter prompts free up memory for a higher-quality quantization.
For most users the best balance is Q4_0, needing about 18.42 GB. That means ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF fits entirely in the VRAM of a 24 GB GPU or larger, running fully on the GPU.
Available GGUF quantizations for ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF include Q4_0, Q8_0, BF16. The model supports a native context length of up to 1,048,576 tokens; a longer context grows the KV cache and the memory needed.
GGUF file size and memory by quantization
Compare real GGUF weight sizes, estimated KV cache and total memory for Q4, Q5, Q8 and every quantization published in this repository.
| Quant. | Bits | Quality | Weights | KV | Total | Speed~ | Verdict |
|---|---|---|---|---|---|---|---|
| Q4_0 | 4.79 | Good | 17.6 GB | 0.02 GB | 18.42 GB | 2.8 t/s | Offload |
| Q8_0 | 8.51 | Excellent | 31.28 GB | 0.02 GB | 32.1 GB | — | Insufficient |
| BF16 | 16.01 | Excellent | 58.84 GB | 0.02 GB | 59.67 GB | — | Insufficient |
KV cache computed from the model's exact architecture. Speed is a rough estimate bounded by memory bandwidth.
Frequently asked questions
What kind of model is ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF?
ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF is a language model with 31.58 billion parameters, based on the nemotron_h_moe architecture. It is released under the other license and distributed as GGUF files for local inference.
Is ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF a Mixture-of-Experts (MoE) model?
Yes. ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF is a Mixture-of-Experts model with 128 experts, of which 6 are activated per token. That makes it faster than a dense model of the same size, but all 128 experts must be loaded into memory, so the VRAM/RAM it needs is driven by the total parameter count, not the active one.
How does ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF handle long context?
ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF uses a hybrid attention design: only 6 of its 52 layers keep a full KV cache, and the rest use linear attention with a constant-size state. Long contexts cost far less memory than on a conventional model of this size.
How much VRAM do you need to run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF?
You need about 18.42 GB of VRAM to run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF entirely on the GPU using the Q4_0 quantization (at a 4,096-token context). Smaller quantizations lower the requirement at the cost of quality.
Can I run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF on an 8 GB GPU?
Partially. ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF only fits on an 8 GB GPU by offloading part of it to system RAM (with Q4_0), which runs but is slower.
Can I run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF on a 16 GB GPU?
Partially. ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF only fits on a 16 GB GPU by offloading part of it to system RAM (with Q8_0), which runs but is slower.
Can I run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF on a 24 GB GPU?
Yes. With 24 GB of VRAM you can run ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF fully on the GPU using Q4_0 (about 18.42 GB).
What context length does ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF support?
ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF supports a native context length of up to 1,048,576 tokens. A longer context grows the KV cache, so it increases the memory needed to run the model.
What is the best quantization for ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF?
For ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF, higher bits-per-weight means better quality. A common sweet spot is a Q4_K_M or Q5_K_M quantization, which keeps most of the quality while roughly halving the memory versus 8-bit. Pick the highest quantization that still fits in your VRAM.