The Cheapest Rig That Runs Nemotron 3.5 Lightning (2026)

Bottom Line (August 11, 2026)

• Cheapest rig that runs it at all: RTX 5060 Ti 16GB + 64GB DDR5, roughly $1,700–$2,400 in parts. The weights do not fit in VRAM, so you offload MoE experts to system RAM. Only 3B parameters are active per token, so this hurts far less than it would on a dense 30B.
• Cheapest rig that runs it well: RTX 5090 32GB. It is the only GeForce card NVIDIA names on the model card, it has native NVFP4, and 32GB holds the 21.6GB checkpoint with real KV headroom. It is also $4,300–$5,000 as of August 2026, which is the problem.
• Cheapest NVIDIA-blessed box: ASUS Ascent GX10 at $3,999 (1TB). Same GB10 silicon as the DGX Spark, 128GB unified memory, runs NVIDIA’s published DSpark recipe, and undercuts the DGX Spark’s $4,699.
• Budget 24GB route: used RTX 3090 at $1,000–$1,300. Ampere is supported through W4A16, but there is no native NVFP4 and only ~2GB of headroom over the weights. Expect a reduced context window.
• Do not buy for the 1M context. Nothing under 32GB serves it. The 1M window is a data-center and GB10 feature.

All prices are US street ranges as of August 2026, during an active DRAM and GDDR7 shortage. MSRPs are not prices right now.

What Actually Shipped Today

Nemotron 3.5 Lightning 30B-A3B is a hybrid: interleaved Mamba-2 layers, MoE layers, and a few Attention layers. 30B total, 3B active. Pre-trained on 20T+ tokens with an NVFP4 recipe, up to 1M context, released under OpenMDW-1.1, which permits commercial use. Model card release date: August 11, 2026.

The scores are strong for the size: MMLU Pro 81.62, GPQA Diamond 75.57, SWE-bench Verified 52.80, IFBench 72.88 — all on the NVFP4 checkpoint, measured by NVIDIA. The NVFP4 numbers track BF16 within about a point, and on SWE-bench Verified NVFP4 actually scores higher (52.80 vs 51.56). That matters for this post: quantizing to NVFP4 costs you almost nothing, so the cheap route is not the compromised route.

The selling point is speed, not peak intelligence. NVIDIA claims up to 4x the output speed of similar-sized models and 86% on PinchBench while finishing 10,000 tasks 30% faster than Qwen3.6 35B. With 3B active, this model is bandwidth-light. That is exactly what a long-running local agent needs.

The Numbers That Decide Your Build

We pulled the actual file sizes rather than trusting the round numbers:

Read the first row again: 21.6GB of weights, plus a 1.2–1.3GB drafter if you want speculative decoding, plus KV cache and Mamba state. That is the whole sizing story. A 24GB card is holding roughly 23GB before it serves a single token. A 32GB card has room to work.

Tiers: Cheapest to Best

Budget: RTX 5060 Ti 16GB + 64GB DDR5 — ~$1,700–2,400

The weights do not fit. Run it anyway.

MoE models tolerate offload in a way dense models do not. Only 3B parameters activate per token, so keeping attention, Mamba layers and the KV cache on the GPU while the sleeping experts sit in system RAM costs you far less throughput than the 21.6GB figure suggests. This is the --n-cpu-moe pattern — see llama.cpp MoE offload flags explained.

The 5060 Ti is the cheapest card with native NVFP4 hardware, so the parts that do live on the GPU run at Blackwell speed. Budget as of August 2026: GPU $589–805, 64GB DDR5-4800 CL40 kit ~$680–880, B650 board $140–215, 650–750W Gold PSU $80–120, 1TB NVMe $90–165 (go 2TB if you plan to keep the BF16 checkpoint too). Add a CPU and case.

Get 64GB, not 32GB. The experts have to live somewhere.

Budget alt: used RTX 3090 24GB — ~$1,700–2,400

NVIDIA lists Ampere via W4A16 on the model card and publishes a working vLLM command for it, serving the same NVFP4 checkpoint through W4A16 kernels. So yes, a 3090 genuinely runs this model — that claim survives checking.

Two honest caveats. Ampere has no native NVFP4 hardware, so W4A16 buys you the memory footprint and not the Blackwell throughput. And 21.6GB of weights in 24GB of VRAM leaves roughly 2GB for everything else, so you will be setting --max-model-len well below 1M and running an FP8 or Q8 KV cache. See KV cache quantization: Q8 vs Q4.

The 3090 is $1,000–1,300 used as of August 2026, not the $650–750 that every older guide still quotes. That price move is why this tier costs about the same as the 16GB build: cheaper RAM, dearer GPU.

Mid and best: GB10 boxes and the 5090

The ASUS Ascent GX10 at $3,999 and the DGX Spark at $4,699 run the same GB10 Grace Blackwell chip with 128GB of unified memory. NVIDIA’s DGX Spark recipe serves the full 1M context with FP8 KV cache and the DSpark drafter. If you want the model as NVIDIA intends it, this is the box, and the GX10 is $700 cheaper for the same silicon.

The RTX 5090 32GB is the only GeForce card in the model card’s supported-hardware list. 32GB fits the checkpoint plus a real KV cache, and it has the memory bandwidth GB10 does not — so single-stream decode is fastest here by a wide margin. At $4,300–5,000 as of August 2026 it costs more than a whole GX10, which is an absurd sentence that is nonetheless true during this shortage. Budget for the power too: the 5090 is a 575W card and NVIDIA specifies a 1000W system PSU for it, so this tier needs a 1000W-or-larger ATX 3.1 unit with a native 12V-2x6 cable — not the 650-750W unit the 16GB build above is sized for. More on that trade in Best local LLM for RTX 5090.

Software Status, Day One

Be careful here. Two different pictures:

• NVIDIA’s own model card documents vLLM (nightly v0.27.1, with --moe-backend marlin on GB10 and --moe-backend humming on H100/Ampere), TensorRT-LLM, and SGLang (lmsysorg/sglang:dev-nemotron3-5-lightning). It does not contain a llama.cpp recipe.
• NVIDIA’s launch blog names llama.cpp, Ollama, LM Studio and Unsloth, and Ollama’s library already has the model live today with 11 tags.

So GGUF works, but it is hours old. The hybrid Mamba-2 + MoE architecture is exactly the kind of thing that gets partial support first and correct support later. If you are buying hardware on a llama.cpp assumption, wait a week and watch the issue tracker. If you are on vLLM, the recipes are published and tested.

One more day-one note: the recommended sampling is temperature 1.0, top_p 0.95, and the parsers are --reasoning-parser nemotron_v3 with --tool-call-parser qwen3_coder. Get those wrong and the model looks broken when it is not.

Why This Model Is Worth a Rig at All

Most 30B models are a compromise. This one is a specialist: 3B active means it decodes fast on modest bandwidth, the 1M context means an agent can hold a whole codebase, and OpenMDW-1.1 means you can use it commercially. NVIDIA’s own framing — “long-running autonomous agents, sub-agent workhorse deployments” — is the honest one. It is not going to out-think a frontier model. It is going to run for eight hours on your desk for the cost of electricity.

If that is what you want, buy for sustained throughput and enough memory, not for peak benchmark scores.

See Also

• Best models to run on NVIDIA DGX Spark — the GB10 model landscape this one joins
• Best local LLM for RTX 5090 — the 32GB consumer tier in detail
• RTX 5090 vs 4090 vs used 3090 — which GPU to actually buy in 2026
• llama.cpp MoE offload flags explained — how the 16GB build works
• KV cache quantization: Q8 vs Q4 — how to claw back context on 24GB
• Best local LLM by GPU (hub) — per-card picks