NVIDIA · Chat and reasoning · 253B parameters · other
Llama 3.1 Nemotron Ultra 253B v1
Needs 304 GB at FP8 / INT8. The smallest real configuration today is 4× A100 80GB at Vast.ai, $2.40 per hour, about $1,754 a month at 730 hours.
Parameters253Bdense
Context128k tokensmaximum published
Checkpoint507 GBBF16, as published
Licenseothercheck before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
4× A100 80GB320 GB VRAM · 95% used
- Per hour
- $2.40
- Per month
- $1,754
- Cheapest at
Vast.ai
Memory · 3 precisions · 4 context lengthsHow much VRAM it needs
Weights plus runtime overhead, then the KV cache that grows with the context you serve.
| Precision | Weights | Weights + headroom | 4k context | 32k context | 128k context |
|---|---|---|---|---|---|
| BF16published checkpoint | 507 GB | 609 GB | 652 GB | 957 GB | 2,000 GB |
| FP8 / INT8estimated | 253 GB | 304 GB | 348 GB | 652 GB | 1,696 GB |
| INT4 / MXFP4estimated | 152 GB | 183 GB | 226 GB | 531 GB | 1,574 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 44 GB | 348 GB | 1,392 GB |
Cross-checked on Hugging Face Oct 28, 2026: 253.4B parameters in the safetensors headers, 506.8 GB of weight files, attention layout from config.json (162 layers, 128 KV heads × 128). Method: weights as published at the native precision, verified on Hugging Face 2026-10-28; other precisions estimated at 1 (8-bit) and 0.6 (4-bit) bytes per parameter; × 1.2 runtime headroom; KV cache = 65536 bytes per token per layer × 162 full-attention layers (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 110 GPUs · cheapest eligible on-demand offer per GPU count, every provider on fileWhat it costs to run, GPU by GPU
The fewest GPUs of one type whose total memory fits the model at short context, priced from the cheapest eligible on-demand offer for exactly that count (45-day window, newest reading per provider), with the next providers beside it. Monthly = 730 hours.
| GPU | Smallest config | VRAM used | $ / hour | $ / month | Cheapest at |
|---|---|---|---|---|---|
| A100 80GBBest fit | 4× 80 GB320 GB total | 95% tight | $2.40 | $1,754 | Vast.aiTrust 78 · read Nov 1 · receipt · also Spheron $4.76, Gcore $5.63 |
| A100 40GB | 8× 40 GB320 GB total | 95% tight | $3.20 | $2,338 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $4.67, Denvr Cloud $9.20 |
| RTX A6000 | 8× 48 GB384 GB total | 79% | $3.36 | $2,453 |  LiumTrust 40 · read Nov 1 · receipt · also LeaderGPU $3.58, GPU.ai $3.76 |
| L40S | 8× 48 GB384 GB total | 79% | $3.74 | $2,727 | Vast.aiTrust 78 · read Nov 1 · receipt · also Lium $6.00, GPU.ai $6.40 |
| RTX PRO 6000 | 4× 96 GB384 GB total | 79% | $3.84 | $2,805 | Vast.aiTrust 78 · read Nov 1 · receipt · also UpCloud $3.89, Hetzner $6.44 |
| Gaudi 2 | 4× 96 GB384 GB total | 79% | $4.18 | $3,052 |  Cyfuture AInot scored · read Nov 1 · receipt |
| A40 | 8× 48 GB384 GB total | 79% | $4.61 | $3,364 |  LeaderGPUTrust 59 · read Nov 1 · receipt · also Exoscale $8.36 |
| A100X | 4× 80 GB320 GB total | 95% tight | $5.12 | $3,739 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 6000 Ada | 8× 48 GB384 GB total | 79% | $5.20 | $3,796 |  GPU.ainot scored · read Nov 1 · receipt · also io.net $5.46, QuantaCloud $6.21 |
| MI300X | 2× 192 GB384 GB total | 79% | $5.71 | $4,167 | Cyfuture AInot scored · read Nov 1 · receipt · also Hot Aisle $5.98, AtmosCompute $6.69 |
| MI325X | 2× 256 GB512 GB total | 59% | $6.21 | $4,536 | Cyfuture AInot scored · read Nov 1 · receipt |
| H100 | 4× 80 GB320 GB total | 95% tight | $6.84 | $4,991 |  GcoreTrust 80 · read Nov 1 · receipt · also Edgevana $7.98, UpCloud $8.00 |
| B200 | 2× 180 GB360 GB total | 84% | $8.22 | $6,002 |  UpCloudTrust 75 · read Nov 1 · receipt · also Deep Infra $9.78, Elice AI Cloud $13.64 |
| L40 | 8× 48 GB384 GB total | 79% | $8.40 | $6,132 |  OblivusTrust 82 · read Nov 1 · receipt · also CoreWeave $10.00, AtmosCompute $12.21 |
| RTX PRO 6000D | 4× 84 GB336 GB total | 90% tight | $10.56 | $7,709 | AtmosComputenot scored · read Nov 1 · receipt |
| H200 NVL | 4× 141 GB564 GB total | 54% | $10.67 | $7,788 | Vast.aiTrust 78 · read Nov 1 · receipt · also IONOS $13.04, Geodd $13.56 |
| H200 | 4× 141 GB564 GB total | 54% | $11.05 | $8,070 | Vast.aiTrust 78 · read Nov 1 · receipt · also Theta EdgeCloud $14.76, Canopy Wave $16.00 |
| B300 | 2× 288 GB576 GB total | 53% | $11.65 | $8,506 | UpCloudTrust 75 · read Nov 1 · receipt · also FPT AI Factory $13.98, Vast.ai $17.00 |
| H100 NVL | 4× 94 GB376 GB total | 81% | $12.48 | $9,112 | Vast.aiTrust 78 · read Nov 1 · receipt · also EmpirioLabs AI $22.40, Open Telekom Cloud $25.90 |
| B200 CC | 2× 180 GB360 GB total | 84% | $18.09 | $13,206 |  VerdaTrust 81 · read Nov 1 · receipt |
| MI355X | 8× 288 GB2304 GB total | 13% | $23.60 | $17,228 |  TensorWaveTrust 64 · read Nov 1 · receipt |
| GB300 | 2× 288 GB576 GB total | 53% | $26.11 | $19,060 | VerdaTrust 81 · read Nov 1 · receipt |
| B300 CC | 4× 288 GB1152 GB total | 26% | $46.31 | $33,806 | VerdaTrust 81 · read Nov 1 · receipt |
| A800 80GB PCIe | 8× 80 GB640 GB total | 47% | $50.80 | $37,088 |  Volcano Enginenot scored · read Nov 1 · receipt |
Throughput estimates · Oct 4, 2026Tokens per second, and what a million costs
Chat and reasoning. Speed estimated from memory bandwidth and model size, never measured. The cost per 1M output tokens is an estimate at the prices read on Oct 4, 2026; it is not recomputed here.
| GPU | Cards | Precision | tok/s per user | per 1M output tokens · estimate |
|---|---|---|---|---|
| MI300X | 4 | FP8 | 24.2 | $7.25 |
| B200 | 4 | FP8 | 46.5 | $9.73 |
| H200 | 4 | FP8 | 30.4 | $12.60 |
| H100 SXM | 8 | FP8 | 35.0 | $13.60 |
| RTX PRO 6000 | 8 | FP8 | 13.8 | $18.40 |
| H100 PCIe | 8 | FP8 | 14.7 | $24.00 |
cards, precision and speed estimated from memory bandwidth and model size; cost recomputed by the site from verified prices · cheapest per output at the prices of Oct 4, 2026: MI300XWhy estimates, and their limits →
Llama 3.1 Nemotron Ultra 253B v1 · related