Nemotron 3.5 Lightning 30B-A3B (MoE)
Nemotron 3.5 Lightning 30B-A3B (MoE) needs 19.7 GB of VRAM at Q4_K_M with 8k context, computed from its config.json as checked on 3 November 2026. The smallest card that holds it has 24 GB, and the fastest desktop card that runs it, the RTX 5090, generates at most 721 tokens per second.
- Parameters
- 31.6B, 3B active
- Context
- 262,144 tokens
- Licence
- OpenMDW 1.1
- Released
- August 2026
- Config checked
- 3 November 2026
Open Nemotron 3.5 Lightning 30B-A3B (MoE) in the LLM checker
Which card size runs it
19.7 GB against each memory size. A model may use 95% of a card, and only what is left over, or missing, is coloured; tight means above 85%, with little room for a longer conversation.
Other quantisations, context lengths or a Q8 KV cache: the VRAM calculator.
Computed 3 November 2026Which GPU to pick
The smallest card here that loads it at Q4_K_M with 8k of context is the RTX 5090 Laptop GPU (24 GB): 19.7 GB needed, an estimated 360 tokens per second. The same card still has room at 32,768 tokens of context (19.8 GB), so there is no reason to buy above it for this model.
At Q8_0, which is near-lossless, the card is the RTX A6000 (48 GB), needing 35.4 GB. On Apple silicon, a Mac mini M4 Pro holds it with 32,768 tokens of context at an estimated 110 tokens per second; prompt processing is slower than on Nvidia, which long documents make obvious. Among unified-memory desktops, the Ryzen AI Max+ 395 holds it at 32,768 tokens in 19.8 GB, at an estimated 103 tokens per second.
Speed on each card
| GPU | Fits | Est. tokens/s | Feels like | Memory | Bandwidth |
|---|---|---|---|---|---|
| H100 SXM | Yes | ≤ 1,348 | faster than you can read | 80 GB | 3,350 GB/s |
| A100 | Yes | ≤ 820 | faster than you can read | 80 GB | 2,039 GB/s |
| RTX 5090 | Yes | ≤ 721 | faster than you can read | 32 GB | 1,792 GB/s |
| RTX PRO 6000 Blackwell | Yes | ≤ 721 | faster than you can read | 96 GB | 1,792 GB/s |
| RTX PRO 5000 Blackwell | Yes | ≤ 541 | faster than you can read | 48 GB | 1,344 GB/s |
| RTX 4090 | Yes | ≤ 406 | faster than you can read | 24 GB | 1,008 GB/s |
| RX 7900 XTX | Yes | ≤ 386 | faster than you can read | 24 GB | 960 GB/s |
| RTX 6000 Ada | Yes | ≤ 386 | faster than you can read | 48 GB | 960 GB/s |
| RTX 3090 | Yes | ≤ 377 | faster than you can read | 24 GB | 936 GB/s |
| RTX 5090 Laptop GPU | Yes | ≤ 360 | faster than you can read | 24 GB | 896 GB/s |
| RTX PRO 4500 Blackwell | Yes | ≤ 360 | faster than you can read | 32 GB | 896 GB/s |
| L40S | Yes | ≤ 348 | faster than you can read | 48 GB | 864 GB/s |
| RTX A6000 | Yes | ≤ 309 | faster than you can read | 48 GB | 768 GB/s |
| RTX PRO 4000 Blackwell | Yes | ≤ 270 | faster than you can read | 24 GB | 672 GB/s |
| Radeon AI PRO R9700 | Yes | ≤ 257 | faster than you can read | 32 GB | 640 GB/s |
| Arc Pro B70 | Yes | ≤ 245 | faster than you can read | 32 GB | 608 GB/s |
| Arc Pro B60 | Yes | ≤ 183 | faster than you can read | 24 GB | 456 GB/s |
| RTX PRO 4000 Blackwell SFF | Yes | ≤ 174 | faster than you can read | 24 GB | 432 GB/s |
| RTX Spark | Yes | ≤ 121 | faster than you can read | 128 GB (102 for the GPU) | 300 GB/s |
| DGX Spark | Yes | ≤ 110 | faster than you can read | 128 GB (126 for the GPU) | 273 GB/s |
| Ryzen AI Max+ 395 | Yes | ≤ 103 | faster than you can read | 128 GB (96 for the GPU) | 256 GB/s |
| RX 7900 XT | No | - | does not fit | 20 GB | 800 GB/s |
| RTX 4060 Ti | No | - | does not fit | 16 GB | 288 GB/s |
| RTX 5060 Ti | No | - | does not fit | 16 GB | 448 GB/s |
| RTX 5080 | No | - | does not fit | 16 GB | 960 GB/s |
| RTX 5070 Ti | No | - | does not fit | 16 GB | 896 GB/s |
| RTX 4080 SUPER | No | - | does not fit | 16 GB | 736 GB/s |
| RX 9070 XT | No | - | does not fit | 16 GB | 640 GB/s |
| RX 9060 XT | No | - | does not fit | 16 GB | 320 GB/s |
| RTX 3060 | No | - | does not fit | 12 GB | 360 GB/s |
| RTX 4070 | No | - | does not fit | 12 GB | 504 GB/s |
| RTX 5070 | No | - | does not fit | 12 GB | 672 GB/s |
| Arc B580 | No | - | does not fit | 12 GB | 456 GB/s |
Single-stream decode, from memory bandwidth at 70% efficiency. Prompt processing and batching not included. These are ceilings, not measurements: Nemotron 3.5 Lightning 30B-A3B (MoE) reads only 1.7 GB of weights per token at Q4, so little that the per-token costs the formula leaves out decide much of the real speed. The 21 figures marked ≤ are where a real runtime falls furthest below the number shown. See the speed estimator for other quantisations and Apple hardware, or every card compared if you are choosing hardware rather than a model.
Computed 3 November 2026Memory by quantisation and context
| Quant | 4,096 ctx | 8,192 ctx | 32,768 ctx |
|---|---|---|---|
| Q8_0 | 35.4 GB · fits 48 GB | 35.4 GB · fits 48 GB | 35.6 GB · fits 48 GB |
| Q5_K_M | 23.9 GB · fits 32 GB | 23.9 GB · fits 32 GB | 24.1 GB · fits 32 GB |
| Q4_K_M | 19.6 GB · fits 24 GB | 19.7 GB · fits 24 GB | 19.8 GB · fits 24 GB |
Weights at this quant: 18.3 GB at Q4. Every extra 1,000 tokens of context adds about 0.006 GB of KV cache at FP16, on top of a fixed 0.05 GB of recurrent state that does not grow. Try other settings in the VRAM calculator.
About the model
Nemotron 3.5 Lightning 30B-A3B (MoE) has 31.6B parameters with 3B active per token (mixture of experts). It has 52 blocks, of which only 6 are attention layers (2 KV heads of dimension 128); 23 are Mamba-2 layers with a small fixed state and 23 are mixture-of-experts layers with no cache at all. NVIDIA's open 30B-A3B model, August 2026, released with its training data and recipes. A three-way hybrid: Mamba-2 layers, mixture-of-experts layers, and only six attention layers, which makes it the cheapest model here to give a long context.
Text only 262,144-token window OpenMDW 1.1 August 2026 3B of 31.6B active
Reasoning. A reasoning and chat model; reasoning output is parsed separately by the runtime.
Native context window: 262,144 tokens. Licence: OpenMDW 1.1, a custom licence with its own conditions; read it before commercial use. Weights published in August 2026.
Good at
- Very long context on a 24 GB card: six attention layers with two KV heads each is almost no cache at all.
- Teams that want to see, and reuse, the training data and recipe, not only the weights.
- Fine-tuning: NVIDIA describes the BF16 release as the starting point for customisation.
Watch out for
- The BF16 checkpoint is meant for customisation. For running it, NVIDIA points to its NVFP4 and GGUF releases.
- OpenMDW 1.1 is its own licence; read it rather than assuming Apache.
- The card lists six supported languages: English, Spanish, French, German, Italian and Japanese.
Similar models
The models that need about the same memory as Nemotron 3.5 Lightning 30B-A3B (MoE) at Q4_K_M and 8k, which makes them the real alternatives on whatever card you have. The context column is what each extra 1,000 tokens costs, and it is where models of the same size differ most.
| Model | Size | Needs | Per 1k context | Window | Released |
|---|---|---|---|---|---|
| Nemotron 3.5 Lightning 30B-A3B (MoE) | 31.6B · 3B active | 19.7 GB | 0.006 GB | 256k | 2026-08 |
| Qwen3 30B-A3B (MoE) | 30.5B · 3.3B active | 19.7 GB | 0.098 GB | 40k | 2025-04 |
| GLM-4.7 Flash 30B-A3B (MoE) | 31.2B · 3B active | 19.8 GB | 0.054 GB | 198k | 2026-01 |
| Granite 4.2 30B | 29.3B | 20.3 GB | 0.262 GB | 128k | 2026-08 |
| Muse Glimmer 30B | 29.8B | 18.7 GB | 0.013 GB | 128k | 2026-08 |
See every model compared, which models fit on 8 to 96 GB, or the cheapest hour of every cloud card.
Questions
- Can I run Nemotron 3.5 Lightning 30B-A3B (MoE) on a 24 GB card?
- Yes, at Q4_K_M and 8k context it needs about 19.7 GB, which fits a 24 GB card with room.
- How much VRAM does Nemotron 3.5 Lightning 30B-A3B (MoE) need at Q8?
- About 35.4 GB at 8k context, or 35.6 GB at 32k. Q8 is near-lossless; use it when it fits.
- What is the context length of Nemotron 3.5 Lightning 30B-A3B (MoE)?
- 262,144 tokens natively. Holding all of it at Q4_K_M takes about 21.2 GB, of which 1.7 GB is context.
- Why does Nemotron 3.5 Lightning 30B-A3B (MoE) need so little memory for long context?
- Because most of its layers do not keep a cache that grows. It has 52 blocks, of which only 6 are attention layers (2 KV heads of dimension 128); 23 are Mamba-2 layers with a small fixed state and 23 are mixture-of-experts layers with no cache at all. At 32k tokens the context costs about 0.3 GB; if all 52 layers kept a conventional full-length cache, the same conversation would cost about 1.7 GB.
- How fast is Nemotron 3.5 Lightning 30B-A3B (MoE) on an RTX 4090?
- At most about 406 tokens per second at Q4, single stream, which is faster than you can read. That is a ceiling worked out from the card's memory bandwidth, not a measurement, and at a figure this high real runtimes land well below it, because the costs the formula leaves out take a growing share of each token.
- Architecture values from config.json in nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 on Hugging Face, and its model card, checked 3 November 2026. Every figure on this page is computed from those values; check them before buying hardware for this model.
- Estimates assume a single conversation on a card that is otherwise free. A desktop on the same GPU takes 0.5 to 2 GB.