Skip to content

Cost reference

What fine-tuning and training cost

Estimates from the same arithmetic as the LLM training cost calculator: FLOPs per token from model size, real GPU throughput at a stated utilisation, and on-demand rental prices. Each page shows its assumptions and sources.

Checked 15 November 2026 · Prices as of November 2026

Fine-tuning an open model

ModelCheapest setupPer 1M tokensPer 100M tokens
Llama 3.1 8BLoRA on L40S$0.067$6.72
Llama 3.1 70BLoRA on 4 × L40S$0.59$59.01
Llama 3.2 3BLoRA on L40S$0.027$2.69
Llama 3.3 70BLoRA on 4 × L40S$0.59$59.01
Qwen3 8BLoRA on L40S$0.069$6.85
Qwen3 14BLoRA on L40S$0.12$12.35
Qwen3 32BLoRA on 2 × L40S$0.27$27.40
Qwen3 30B-A3BLoRA on 2 × L40S$0.028$2.76
Qwen2.5 7BLoRA on L40S$0.064$6.37
Mistral 7BLoRA on L40S$0.061$6.06
Gemma 3 27BLoRA on 2 × L40S$0.23$22.94
gpt-oss-20bLoRA on 2 × L40S$0.030$3.01

Pretraining from scratch

Model sizeTokens (20 per parameter)H100-hoursCost on H100s
125M2.5B1.3$5.78
1B20B84$370
3B60B758$3,329
7B140B4,129$18,125
13B260B14,240$62,513
70B1,400B412,875$1.8M

More

Sources

  1. [1]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  2. [2]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  3. [3]Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
  4. [4]PyTorch. torchtune: memory and throughput benchmarks (Llama 3.1 8B on RTX 4090)
  5. [5]Together AI. Fine-tuning pricing (seen 15 November 2026)
  6. [6]OpenAI. API pricing: fine-tuning (seen 15 November 2026)
  7. [7]Runpod. GPU pricing (updated 12 November 2026)
  8. [8]Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
  9. [9]Hoffmann et al., 2022. Training Compute-Optimal Large Language Models (arXiv:2203.15556)
  10. [10]Meta, 2024. The Llama 3 Herd of Models (arXiv:2407.21783)
  11. [11]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)

Measure before you spend on GPUs.

Kojen runs prompt tuning and model search for real against your own test cases, so you know whether training is worth it.

Start training free