Lowest-cost LoRA fine-tuning solution: Train your own model for under ¥10

TL;DR — Straight-to-the-point answer：QLoRA fine-tuning of a 7B model,Renting one 24GB RTX 3090/4090 for 3–4 hours on a pay-as-you-go basis is sufficient—total cost ¥5–10Three core strategies: load models in 4-bit, train LoRA adapters only, use spot/preemptible instances + checkpointing.

Why LoRA Reduces Costs

Full fine-tuning or pretraining a 7B model requires ≥60GB VRAM (model + Adam optimizer states + gradients), so only A100 rentals apply—at ¥6–12/hour. LoRA freezes the base model and trains only inserted low-rank matrices (~1% of parameters); QLoRA further loads the base model in 4-bit quantized form:

• Model weights in 4-bit: ~4–5 GB
• LoRA parameters + optimizer: ~2 GB
• Activation memory (batch 8, seq 2048): ~8–12 GB
• Total 16–20 GB — a consumer-grade 24 GB GPU handles it perfectly

GPU Model Selection & Real-Time Pricing

The 3090 is the cost killer; the 4090 is the speed champion. Both work — choose based on your budget.

End-to-End Cost Control (using Qwen2.5-7B as an example)

1. Select instance: RTX 3090 pay-as-you-go; spot/idle instances prioritized (30% cheaper, interruptible)
2. Setup environment in 10 minutes：Use transformers + peft + bitsandbytes + trl; save time by selecting prebuilt LLM images from the marketplace
3. DataThousands of instruction samples are sufficient; prepare data in JSON format and upload to cloud storage.
4. Training takes 2–3 hours：LoRA rank 16, batch size 4 + gradient accumulation 4, train for 2–3 epochs;Save checkpoint every 500 steps(Prevent spot instance termination)
5. Export & merge：Merge LoRA weights into the base model, then download locally or push to HuggingFace

Total ≈ 4 hours × ¥1.4/hour ≈ ¥5.6, Including upload/download time, total cost stays under ¥8.

Three Common Pitfalls

• Out of memory? Reduce sequence length first：Reducing sequence length from 2048 to 1024 cuts activation memory usage by half—more cost-effective than lowering batch size
• Always enable auto-save for spot instancesProgress isn’t lost upon preemption—simply restart and resume training from the latest checkpoint.
• Don’t use free-tier credits for production tasksFree trial credits from various platforms are ideal for environment testing; use top-up balance for formal training to avoid service interruption.

Prices across platforms change hourly; the table above shows real-time price comparison at this moment (updated hourly). Unsure whether GPU memory is sufficient? TryPickerCalculate costs before ordering. Some links contain referral commissions, but they do not affect ranking.