Written by Jakub Rusinowski · Last updated September 11, 2026
Model library → Nemotron Cascade 2 → Nemotron-Cascade 2 30B-A3B
31.6B parameters resident — about 19.9 GB at Q4_K_M, so a 24 GB card holds it — with 3.5B active per token, which is why it decodes far faster than its size suggests. 52 layers: 23 Mamba-2 state-space, 23 MoE, and 6 GQA attention. Each MoE layer carries 128 routed experts plus one shared, six routed activated per token. Gold-medal level at IMO 2025, IOI and the ICPC World Finals, reached through post-training rather than scale — and NVIDIA published that recipe. 1M-token context, NVIDIA Open Model License.
Nemotron-Cascade 2 30B-A3B needs about 20 GB of VRAM at Q4_K_M — quantized weights plus framework overhead, before any KV cache. On Apple Silicon that figure comes out of unified memory.
| Parameters | 31.6 Billion (3.5B active) |
| Context window | 1,048,576 |
| Architecture | Hybrid Mamba-2 + MoE + GQA attention |
| Provider | NVIDIA |
| Licence | NVIDIA Open Model License |
| Specified at | Q4_K_M |
| System RAM | 32 GB |
| Record updated | 2026-09-11 |
NVIDIA Open Model License — commercial use permitted. Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Modelled on a reference NVIDIA RTX 4090 (24 GB), with no KV cache (this record has no published architecture). Speed figures are ESTIMATES from the memory-bandwidth roofline described on the methodology page, not benchmarks we ran — rows marked measured come from published or reader-submitted runs. VRAM here includes the KV cache, so it reads higher than the headline figure above, which does not.
| Quant | Weights | VRAM needed | Est. speed | Fit on 24 GB |
|---|---|---|---|---|
| Q2_K | 10.4 GB | 11.2 GB | ~201 tok/s (est.) | Fits comfortably |
| Q3_K_M | 13.5 GB | 14.3 GB | ~185 tok/s (est.) | Fits comfortably |
| Q4_K_M | 19.1 GB | 19.9 GB | ~160 tok/s (est.) | Fits comfortably |
| Q5_K_M | 22.4 GB | 23.2 GB | ~149 tok/s (est.) | Tight fit |
| Q6_K | 25.9 GB | 26.7 GB | ~21 tok/s (est.) | Offloads to system RAM (slow) |
| Q8_0 | 33.6 GB | 34.4 GB | ~18 tok/s (est.) | Offloads to system RAM (slow) |
| F16 | 63.2 GB | 64.0 GB | — | Won't fit |
Want the memory numbers alone, at every quantization level and your own context length? Use the Nemotron-Cascade 2 30B-A3B VRAM calculator.
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.
The cheapest catalogued GPU that runs Nemotron-Cascade 2 30B-A3B is the AMD Radeon RX 7900 XT (20 GB).
Install Ollama, then run:
ollama run nemotron-cascade-2:30bWeights on Hugging Face: nvidia/Nemotron-Cascade-2-30B-A3B.
Best for: reasoning, agentic tasks, math, long context.
← All Nemotron Cascade 2 models | VRAM calculator | Build a PC for this model | Check your own hardware