NVIDIA · Chat and reasoning · 49.9B parameters · other
Llama 3.3 Nemotron Super 49B v1.5
Needs 60 GB at FP8 / INT8. The smallest real configuration today is 2× Tesla V100 32GB at Vast.ai, $0.14 per hour, about $99 a month at 730 hours.
Parameters49.9Bdense
Context128k tokensmaximum published
Checkpoint100 GBBF16, as published
Licenseothercheck before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
2× Tesla V100 32GB64 GB VRAM · 94% used
- Per hour
- $0.14
- Per month
- $99
- 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 | 100 GB | 120 GB | 131 GB | 206 GB | 464 GB |
| FP8 / INT8estimated | 50 GB | 60 GB | 71 GB | 146 GB | 404 GB |
| INT4 / MXFP4estimated | 30 GB | 36 GB | 47 GB | 122 GB | 380 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 11 GB | 86 GB | 344 GB |
Cross-checked on Hugging Face Oct 28, 2026: 49.9B parameters in the safetensors headers, 99.7 GB of weight files, attention layout from config.json (80 layers, 64 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 = 32768 bytes per token per layer × 80 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 |
|---|---|---|---|---|---|
| Tesla V100 32GBBest fit | 2× 32 GB64 GB total | 94% tight | $0.14 | $99 | Vast.aiTrust 78 · read Nov 1 · receipt · also Cyfuture AI $1.13, io.net $2.52 |
| T4 | 4× 16 GB64 GB total | 94% tight | $0.16 | $117 | Vast.aiTrust 78 · read Nov 1 · receipt · also iwinv $0.78, NHN Cloud $2.46 |
| Tesla P100 | 4× 16 GB64 GB total | 94% tight | $0.19 | $138 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.37 |
| RTX 3090 | 4× 24 GB96 GB total | 62% | $0.25 | $183 |  Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $0.58, Lium $0.88 |
| RTX 5060 Ti | 4× 16 GB64 GB total | 94% tight | $0.35 | $258 | Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $0.38, Edgevana $0.68 |
| RTX A4000 | 4× 16 GB64 GB total | 94% tight | $0.36 | $266 | Vast.aiTrust 78 · read Nov 1 · receipt · also Spheron $3.84 |
| RTX A6000 | 2× 48 GB96 GB total | 62% | $0.38 | $274 | Clore.aiTrust 46 · read Nov 1 · receipt · also Lium $0.84, Spheron $0.89 |
| Tesla V100 16GB | 4× 16 GB64 GB total | 94% tight | $0.38 | $275 | Vast.aiTrust 78 · read Nov 1 · receipt · also CloudRift $1.00, AtmosCompute $1.30 |
| Quadro RTX 5000 | 4× 16 GB64 GB total | 94% tight | $0.40 | $294 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3060 | 8× 12 GB96 GB total | 62% | $0.43 | $312 | Vast.aiTrust 78 · read Nov 1 · receipt |
| A800 80GB PCIe | 1× 80 GB80 GB total | 75% | $0.47 | $342 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 3070 | 8× 8 GB64 GB total | 94% tight | $0.50 | $364 | Clore.aiTrust 46 · read Nov 1 · receipt · also Vast.ai $5.23 |
| Quadro RTX 8000 | 2× 48 GB96 GB total | 62% | $0.51 | $371 | Vast.aiTrust 78 · read Nov 1 · receipt |
| L40 | 2× 48 GB96 GB total | 62% | $0.54 | $391 | Vast.aiTrust 78 · read Nov 1 · receipt · also io.net $1.32, GPU.ai $1.56 |
| RTX 2080 Ti | 8× 11 GB88 GB total | 68% | $0.59 | $432 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $0.68 |
| A100 80GB | 1× 80 GB80 GB total | 75% | $0.60 | $440 | Vast.aiTrust 78 · read Nov 1 · receipt · also packet.ai $0.89, GPUBank $0.95 |
| RTX 5060 | 8× 8 GB64 GB total | 94% tight | $0.65 | $476 | Clore.aiTrust 46 · read Nov 1 · receipt |
| RTX PRO 6000 | 1× 96 GB96 GB total | 62% | $0.66 | $482 |  packet.ainot scored · read Nov 1 · receipt · also Runpod $0.69, HyperAI $0.80 |
| Titan RTX | 4× 24 GB96 GB total | 62% | $0.70 | $508 | Vast.aiTrust 78 · read Nov 1 · receipt |
| RTX 5070 Ti | 4× 16 GB64 GB total | 94% tight | $0.70 | $508 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.80, Edgevana $0.80 |
| RTX 4070 Ti Super | 4× 16 GB64 GB total | 94% tight | $0.75 | $548 | Clore.aiTrust 46 · read Nov 1 · receipt |
| L40S | 2× 48 GB96 GB total | 62% | $0.76 | $555 |  LiumTrust 40 · read Nov 1 · receipt · also Vast.ai $1.28, VoltageGPU $1.52 |
| A100 40GB | 2× 40 GB80 GB total | 75% | $0.77 | $566 | Vast.aiTrust 78 · read Nov 1 · receipt · also Edgevana $1.29, AtmosCompute $1.45 |
| RTX 5090 | 2× 32 GB64 GB total | 94% tight | $0.80 | $586 | Vast.aiTrust 78 · read Nov 1 · receipt · also Clore.ai $0.96, Edgevana $1.32 |
| The 24 cheapest of 70 GPU types that fit. Every GPU, any precision and context, in the finder | |||||
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 | 1 | FP8 | 46.0 | $1.06 |
| B200 | 1 | FP8 | 82.4 | $1.47 |
| H200 | 1 | FP8 | 52.7 | $1.84 |
| H100 SXM | 2 | FP8 | 57.3 | $1.99 |
| RTX 5090 | 4 | FP8 | 40.6 | $2.14 |
| RTX 4090 | 4 | FP8 | 26.1 | $2.52 |
| H100 PCIe | 2 | FP8 | 30.2 | $2.71 |
| RTX PRO 6000 | 1 | FP8 | 22.1 | $3.27 |
| A100 80GB | 2 | BF16 | 22.7 | $4.56 |
| L40S | 2 | FP8 | 15.6 | $4.87 |
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 3.3 Nemotron Super 49B v1.5 · related