Can you run Nemotron 3 Ultra Math RL (550B A55B) locally?
Nemotron 3 family · 2026 · 560.5B parameters (55B activated per token) · Hugging Face model card
Nemotron 3 Ultra Math RL is a math-reasoning post-training of NVIDIA's Nemotron 3 Ultra and a mixture-of-experts model with 55B activated parameters per token (official card: Number of model parameters 550B Total / 55B Active; architecture Mamba2-Transformer Hybrid Latent Mixture of Experts with Multi-Token Prediction; config num_experts_per_tok=22 of 512 routed + 1 shared). HF safetensors.total=560,524,578,816 for this math RL checkpoint. Native context is 262,144 tokens by config.json (card: defaults to 256K; up to 1M with VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and --max-model-len 1048576), license other (NVIDIA Open Model License family). Local VRAM fit uses the 55B active count: roughly 48 GB at Q4-class (calculated). Weights still occupy ~550B-class disk, so multi-GPU serving is required. Measured speeds: not yet published here.
Minimum: 48 GB+ memory (Q4-class on 55B active, calculated) · Recommended: 64 GB+ for comfortable context headroom (Q4-class on 55B active, calculated)
How much VRAM does Nemotron 3 Ultra Math RL (550B A55B) need at each quantization?
Nemotron 3 Ultra Math RL (550B A55B) needs 370.4 GB of VRAM at Q4_K_M. The table below lists weights-only size and total VRAM including overhead for each common quantization level with 4k and 32k token contexts.
| Quantization | Bits / weight | Weights only | Total + KV @4k ctx | Total + KV @32k ctx |
|---|---|---|---|---|
| Q4_K_M | 4.8 | 336.3 GB | 370.4 GB | 373.6 GB |
| Q5_K_M | 5.7 | 399.4 GB | 439.8 GB | 442.9 GB |
| Q6_K | 6.6 | 462.4 GB | 509.1 GB | 512.3 GB |
| Q8_0 | 8.5 | 595.6 GB | 655.6 GB | 658.7 GB |
| FP16 | 16 | 1,121.0 GB | 1,233.6 GB | 1,236.8 GB |
Method: weights = parameters × bits-per-weight (Q4_K_M ≈ 4.8, Q5_K_M ≈ 5.7, Q6_K ≈ 6.6, Q8_0 ≈ 8.5, FP16 = 16), plus 10% loading overhead, plus KV cache from the verified architecture config (108 layers, 2 KV heads, 128 head dim). Source: HF-verified 2026-09-30 via api (gated=false params=560.52B safetensors.total=560524578816 ctx=262144 license=other repo=nvidia/Nemotron-3-Labs-Ultra-Math-RL@6849d3ecccbd63f3527d56072a3e9717ec4738c9 layers=108 kv_heads=2). MoE: num_experts=512 num_experts_per_tok=22 active_params_b=55.0 (official model card (550B total, 55B active); config 512 routed experts, 22 per token + 1 shared). VRAM figures are documented calculations, not measurements. No tok/s invented.
Which GPUs can run Nemotron 3 Ultra Math RL (550B A55B) locally?
At Q4_K_M with a 4k context, Nemotron 3 Ultra Math RL (550B A55B) needs 370.4 GB of VRAM. The lists below are computed live from our GPU database and grouped by how much headroom the card has. Disclosure: CompareAIHardware.com participates in the Amazon Associates program and earns from qualifying purchases through links on this page. Affiliate relationships do not influence our recommendations.
Needs 2+ GPUs to run Nemotron 3 Ultra Math RL (550B A55B)
One of these cards is too small on its own, but a pair (tensor or pipeline parallel, ~90% efficiency) covers the 370.4 GB requirement. See our multi-GPU guide for setup.
| GPU | VRAM | Bandwidth | Type | |
|---|---|---|---|---|
| AMD Instinct MI325X | 256 GB ×2 | 6000 GB/s | Data Center GPU | |
| AMD Instinct MI355X | 288 GB ×2 | 8000 GB/s | Data Center GPU | |
| NVIDIA B300 (Blackwell Ultra) | 288 GB ×2 | 8000 GB/s | Data Center GPU |
Can a Mac run Nemotron 3 Ultra Math RL (550B A55B)?
Yes — these Apple Silicon machines fit Nemotron 3 Ultra Math RL (550B A55B) at Q4_K_M, since macOS lets the GPU use about 75% of unified memory. Generation speed is bound by memory bandwidth, so the GB/s column matters as much as capacity.
| Apple Silicon | Unified memory | Usable by GPU (~75%) | Bandwidth |
|---|---|---|---|
| M5 Ultra (Mac Studio) | 512 GB | 384 GB | 1200 GB/s |