What it costs to fine-tune Qwen3 32B
The cheapest setup we compared is LoRA on 2 × L40S: about $0.27 per million training tokens, so $2.74 for 10 million and $27.40 for 100 million. Full fine-tuning costs about $4.85 per 10 million tokens on 4 × B200.
GPU-hours and cost by GPU and method
Per epoch over the token count shown; multiply by your number of epochs. GPU-hours are the same however many GPUs the job is split over.
| GPU · method | GPUs for memory | Tokens/s per GPU | GPU-hours per 1M | 1.0M tokens | 10M tokens | 100M tokens |
|---|---|---|---|---|---|---|
| H100 SXM · LoRA$4.39/hr | 2 | 3,019 | 0.09 | $0.40 | $4.04 | $40.40 |
| H100 SXM · QLoRA$4.39/hr | 1 | 2,355 | 0.12 | $0.52 | $5.18 | $51.79 |
| H100 SXM · Full fine-tuning$4.39/hr | 8 | 2,012 | 0.14 | $0.61 | $6.06 | $60.59 |
| B200 · LoRA$7.99/hr | 1 | 6,868 | 0.04 | $0.32 | $3.23 | $32.32 |
| B200 · QLoRA$7.99/hr | 1 | 5,357 | 0.05 | $0.41 | $4.14 | $41.43 |
| B200 · Full fine-tuning$7.99/hr | 4 | 4,578 | 0.06 | $0.48 | $4.85 | $48.48 |
| L40S · LoRA$1.09/hr | 2 | 1,105 | 0.25 | $0.27 | $2.74 | $27.40 |
| L40S · QLoRA$1.09/hr | 1 | 862 | 0.32 | $0.35 | $3.51 | $35.13 |
| L40S · Full fine-tuning$1.09/hr | 13 (over 8) | 737 | 0.38 | — | — | — |
| RTX 5090 · LoRA$1.39/hr | Doesn't fit | 639 | 0.43 | — | — | — |
| RTX 5090 · QLoRA$1.39/hr | 1 | 499 | 0.56 | $0.77 | $7.74 | $77.41 |
| RTX 5090 · Full fine-tuning$1.39/hr | Doesn't fit | 426 | 0.65 | — | — | — |
Compared with fine-tuning APIs
| Option | Per 1M tokens | 10M tokens | 100M tokens |
|---|---|---|---|
| Rented GPU: LoRA on L40S (our estimate) | $0.27 | $2.74 | $27.40 |
| Together AI | Qwen3 32B not listed | — | — |
| OpenAI gpt-4.1-mini fine-tuning (closed model; closed to new users) | $5.00 | $50.00 | $500 |
API prices are per training token processed (dataset tokens × epochs), seen on 15 November 2026. Together AI charges a minimum per job. OpenAI's pricing page says fine-tuning is closed to new users. Rented GPU estimates leave out setup, failed runs and idle time, which an API absorbs.
Your run
Pre-filled for Qwen3 32B: 10M tokens, 3 epochs, LoRA on an H100.
$12.12
2.8
1.4 hours
3,019
30M tokens processed. Needs 2 × H100 SXM for memory at 2,048-token sequences. GPU-hours assume perfect scaling across GPUs.
Assumptions
- Compute per token: 6 FLOPs per active parameter for full fine-tuning (forward plus backward) and about 4 for LoRA and QLoRA, which skip the frozen weights' gradients. Qwen3 32B uses 32.762B parameters per token.
- Utilisation: 40% of each GPU's dense BF16 peak. torchtune measured 3,083 tokens a second for Llama 3.1 8B LoRA on one RTX 4090, about 60% on the same accounting, so well-tuned code can beat these figures.
- QLoRA runs at 78% of LoRA's speed, the ratio torchtune measured (2,413 vs 3,083 tokens a second), because the 4-bit weights are dequantised on the fly.
- Memory decides the GPU count, from our Qwen3 32B VRAM estimates at 2,048-token sequences. Consumer cards are counted as single GPUs only.
- Prices: on-demand Runpod Secure Cloud rates (page dated 12 November 2026). Community Cloud and spot capacity cost less.
Frequently asked questions
- How much does it cost to fine-tune Qwen3 32B?
- About $0.27 per million training tokens at the cheapest setting we compared (LoRA on 2 × L40S, $1.09/hour per GPU), so about $2.74 for 10M tokens and $27.40 for 100M. Budget for several runs: the first rarely wins.
- How long does it take to fine-tune Qwen3 32B on an H100?
- LoRA processes about 3,019 tokens a second on one H100 at 40% utilisation, so 10M tokens take about 55 minutes of GPU time across 2 H100s. Full fine-tuning needs 8 H100s for memory and about 1.5× the compute.
- Is it cheaper to fine-tune Qwen3 32B through an API?
- Together AI's November 2026 price list doesn't include Qwen3 32B, so the choice is rented GPUs or a hosted model. Rented GPU time costs about $2.74 per 10M tokens by our estimate, plus setup and idle time.
- Should I use LoRA or full fine-tuning for Qwen3 32B?
- Start with LoRA or QLoRA. Full fine-tuning Qwen3 32B needs about 529 GB, so 8 × 80 GB GPUs with sharded optimiser state, and about 1.5× the compute per token. LoRA usually gets close for format, style and task adaptation; move to full fine-tuning only when evaluation shows LoRA has plateaued.
Related
- Qwen3 32B VRAM requirementsWhich GPUs fit each method
- Qwen3 32B on RTX 4090Max settings that fit
- Qwen3 32B on RTX 5090Max settings that fit
- Qwen3 32B on L40SMax settings that fit
- Fine-tune Llama 3.1 8B: costMeta
- Fine-tune Llama 3.1 70B: costMeta
- Fine-tune Llama 3.2 3B: costMeta
- Fine-tune Llama 3.3 70B: costMeta
- Fine-tune Qwen3 8B: costAlibaba Qwen
- What is LoRA?Glossary
Sources
- Hugging Face. Qwen/Qwen3-32B config.json
- Alibaba Qwen. Qwen/Qwen3-32B model card
- Alibaba Qwen. Qwen3 Technical Report (arXiv:2505.09388)
- 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)
Try the cheaper levers first.
Prompt tuning and model search in Kojen cost API calls, not GPU-hours, and often close the gap before any fine-tuning.
Start training free