What are the best budget cloud providers for fine-tuning open-source AI models?

As of November 6, 2026: Together AI's managed fine-tuning for the lowest total cost on supported models (from $0.34 per million training tokens for 8–9B models, $4 minimum per job, so a 30-million-token LoRA run on Llama 3.1 8B costs about $10); RunPod for cheap rented GPUs billed per second (RTX 4090 $0.89/h, A100 80GB $1.79/h, H100 PCIe $2.89/h); Vast.ai for the lowest marketplace prices if you can checkpoint (interruptible instances are 50%+ cheaper); Modal for short or spiky jobs with $30 of free credit a month (H100 about $3.95/h, billed per second); and Together's preemptible H100s at $1.99/h for multi-GPU runs.

How much does it cost to fine-tune an open-source LLM in 2026?

For a LoRA fine-tune of a 7–9B model, usually $5–$30. Together AI bills training tokens (dataset size times epochs): 10 million tokens over 3 epochs on Llama 3.1 8B is 30 million tokens at $0.34, about $10.20. Doing it yourself with QLoRA on a single rented 24–48 GB GPU takes a few hours at $0.50–$1.80 an hour. Full fine-tuning of 70B-class models needs multi-GPU H100 or H200 nodes and runs into hundreds or thousands of dollars.

Is it cheaper to rent GPUs or use a managed fine-tuning API?

A managed per-token API is cheaper for small and medium jobs on models the provider supports, because you pay nothing for setup, failed runs or idle time. Renting GPUs is cheaper when you train often, need a model or method the API does not support, or can use interruptible capacity. Rented GPUs also mean you handle drivers, checkpoints and data transfer.

Which GPU do I need to fine-tune a 7B or 8B model?

With QLoRA (4-bit base model plus LoRA adapters), a single 24 GB card such as an RTX 4090 or 3090 is enough for most 7–8B fine-tunes at modest sequence lengths. A 48 GB card (L40S, RTX A6000) gives headroom for longer sequences. Plain LoRA in 16-bit or full fine-tuning needs an 80 GB A100 or H100, and 70B-class models need several of them.

Quick Answer

Best Budget Cloud Providers for Fine-Tuning Open Models 2026

Published: November 6, 2026

The short answer

For most budget fine-tunes, a per-token managed job beats renting a GPU. Pick by job size and how often you train. As of November 6, 2026:

Together AI Fine-Tuning — cheapest total cost for supported models; pay per training token.
RunPod — cheapest simple GPU rental, per-second billing, consumer and datacenter cards.
Vast.ai — lowest marketplace prices, especially interruptible instances.
Modal — best for short or bursty jobs in Python; $30 free credit monthly.
Together preemptible GPU clusters — cheapest H100s for multi-GPU runs ($1.99/h).
Lambda — simple dedicated instances and 1-Click Clusters for bigger runs.

The comparison

Provider	How you pay	Cheapest useful option	H100 / A100 80GB	Free credit	Best for
Together AI Fine-Tuning	Per training token, minimum per job	$0.34 per 1M tokens (Llama 3.1 8B, Qwen3.5 9B), $4 minimum	— (managed)	—	One-off LoRA/SFT/DPO on supported models
RunPod Pods	Per second	RTX 3090 $0.50/h · RTX 4090 $0.89/h · L40S $1.09/h	$2.89 (PCIe) · $3.99 (SXM) / $1.79	—	DIY QLoRA on one GPU
Vast.ai	Per second, market price	Interruptible instances 50%+ cheaper than on-demand	Market rate	—	Checkpointed jobs, lowest price
Modal	Per second, scale to zero	L40S ≈ $1.95/h	≈ $3.95 / ≈ $2.50	$30/month (Starter)	Short scripted jobs, no idle cost
Together GPU Clusters	Per hour	H100 preemptible $1.99/h	$3.99 on-demand	—	Multi-GPU runs you can restart
Lambda	Per hour	A100 40GB $1.99/h	$4.59 / $3.19 (from Nov 16, 2026)	—	Dedicated 8-GPU nodes, clusters

Prices are list prices read from each pricing page on November 6, 2026; Modal hourly figures are its per-second rates × 3,600. Lambda’s listed prices take effect November 16, 2026.

1. Together AI Fine-Tuning — cheapest total cost

Together charges for training tokens: dataset size × epochs, plus evaluation tokens, with a minimum per job. It supports LoRA and full fine-tuning, supervised (SFT) and preference (DPO) training. Sample SFT prices per million tokens: Llama 3.1 8B, Qwen3.5 9B and Qwen3.5 4B $0.34, gpt-oss-20B $0.40, Qwen3.8 27B and Gemma 4 31B $1.05, Llama 3.3 70B $2.03, gpt-oss-120B $2.50; DPO costs about 2.5× SFT. Minimums are $4 for most models and up to $60 for the largest (Kimi K2.6, GLM-5.2).

Worked example: 10 million tokens of chat data × 3 epochs on Llama 3.1 8B = 30M tokens × $0.34 = about $10. No GPU bill, no failed-setup hours. You then serve the adapter on Together or download it.

Pick it if your base model is on Together’s list and you do not need a custom training loop.

2. RunPod — cheapest straightforward GPU rental

RunPod rents GPUs by the second across 30+ regions. For QLoRA on 7–8B models a 24 GB card is enough: RTX 3090 $0.50/h, RTX 4090 $0.89/h, RTX 5090 (32 GB) $1.19/h. For more headroom, 48 GB cards cost $0.59–$1.09/h (RTX A6000, L40S), A100 80GB $1.79/h, H100 PCIe $2.89/h, H100 SXM $3.99/h, H200 $5.29/h and B200 $7.99/h. Clusters cover multi-node jobs.

Pick it if you run your own training script (Unsloth, Axolotl, TRL) and want a cheap single GPU.

3. Vast.ai — lowest prices if you can be interrupted

Vast.ai is a marketplace: prices are set by supply and demand across 40+ data centres and 68+ GPU types, billed per second with no minimum. Interruptible instances are “50%+ cheaper” but can be reclaimed, so save checkpoints every few hundred steps. Reserved 1-, 3- or 6-month terms cut up to 50%. Hosts vary in reliability and network speed; check the host’s rating and bandwidth before uploading a large dataset.

Pick it if cost matters more than predictability and your trainer resumes from checkpoints.

Modal bills per second and scales to zero, so a 40-minute fine-tune costs 40 minutes. Per-second list rates translate to about $1.95/h for an L40S, $2.50/h for an A100 80GB, $3.95/h for an H100 and $6.25/h for a B200, with up to 8 GPUs per container. The free Starter plan includes $30 of compute a month and 10 concurrent GPUs. Region selection costs 1.15–1.75× and non-preemptible execution 3×, so leave both at defaults for training. You write the job as a Python function, which suits scheduled re-training.

Pick it if you fine-tune occasionally from code and hate paying for idle GPUs.

5. Together GPU Clusters and Lambda — when one GPU is not enough

For multi-GPU or 70B-class runs, Together’s GPU Clusters list H100 at $1.99/h preemptible or $3.99/h on-demand, H200 $2.99 / $5.99 and B200 $4.09 / $8.19, with reserved terms down to $3.19 for H100 (181+ days). Lambda’s instances, from November 16, 2026, list H100 $4.59/h, A100 80GB $3.19/h, A100 40GB $1.99/h and B200 $7.69/h per GPU, and 1-Click Clusters of 16 to 2,000+ GPUs.

How to keep the bill small

Use QLoRA or LoRA, not full fine-tuning, unless evaluation shows LoRA falls short. Unsloth (Apache-2.0, updated November 2026) cuts memory so 8B models fit on 24 GB cards.
Train on less, better data. Tokens × epochs is the bill on Together and roughly the time on rented GPUs. Deduplicate and cut boilerplate first.
Run a 1% pilot to find the right learning rate before the full run.
Checkpoint on interruptible GPUs and push adapters to storage; idle disks keep billing on some providers.
Count serving too. A cheap fine-tune that needs a dedicated endpoint can cost more per month than the training. See best inference providers for LoRA fine-tuned models.

How to choose

Is your model on Together’s fine-tuning list and SFT/DPO enough? Together AI.
Need your own script on one GPU? RunPod (on-demand) or Vast.ai (interruptible).
Occasional jobs from Python? Modal.
Multi-GPU or 70B-class? Together preemptible clusters or Lambda.

Last verified: November 6, 2026.