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.
Fine-tuning an open model
| Model | Cheapest setup | Per 1M tokens | Per 100M tokens |
|---|---|---|---|
| Llama 3.1 8B | LoRA on L40S | $0.067 | $6.72 |
| Llama 3.1 70B | LoRA on 4 × L40S | $0.59 | $59.01 |
| Llama 3.2 3B | LoRA on L40S | $0.027 | $2.69 |
| Llama 3.3 70B | LoRA on 4 × L40S | $0.59 | $59.01 |
| Qwen3 8B | LoRA on L40S | $0.069 | $6.85 |
| Qwen3 14B | LoRA on L40S | $0.12 | $12.35 |
| Qwen3 32B | LoRA on 2 × L40S | $0.27 | $27.40 |
| Qwen3 30B-A3B | LoRA on 2 × L40S | $0.028 | $2.76 |
| Qwen2.5 7B | LoRA on L40S | $0.064 | $6.37 |
| Mistral 7B | LoRA on L40S | $0.061 | $6.06 |
| Gemma 3 27B | LoRA on 2 × L40S | $0.23 | $22.94 |
| gpt-oss-20b | LoRA on 2 × L40S | $0.030 | $3.01 |
Pretraining from scratch
More
- VRAM requirements by modelWhich GPUs fit each method
- Fine-tuning VRAM calculatorAny model size
- What it really costs to train an LLMBlog
- The Economics of Model TrainingFree Academy course
Sources
- Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
- Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
- Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
- PyTorch. torchtune: memory and throughput benchmarks (Llama 3.1 8B on RTX 4090)
- Together AI. Fine-tuning pricing (seen 15 November 2026)
- OpenAI. API pricing: fine-tuning (seen 15 November 2026)
- Runpod. GPU pricing (updated 12 November 2026)
- Kaplan et al., 2020. Scaling Laws for Neural Language Models (arXiv:2001.08361)
- Hoffmann et al., 2022. Training Compute-Optimal Large Language Models (arXiv:2203.15556)
- Meta, 2024. The Llama 3 Herd of Models (arXiv:2407.21783)
- 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