Train

Managed LoRA / fine-tune services on top of trainable models, plus self-host on rented H100s. Pick by provider, model, or cost.

est. wall-clock 0.28h
provider

managed training services · 13

cheapest first · click row → service
modelvendorprovidertype$/1k stepsest. total (1.0k)notes
wan-2.7-image-proAlibabaself-host-vastLoRA$0.50$0.50
wan-2.7-imageAlibabaself-host-vastLoRA$0.50$0.50
qwen-image-2.0Alibabaself-host-vastLoRA$0.50$0.50
flux-2-klein-4bBFLself-host-vastLoRA$0.50$0.50Apache 2.0 — fully commercial
qwen-image-2.0-proAlibabafalLoRA$0.95$0.95
qwen-image-2.0AlibabafalLoRA$0.95$0.95
flux-2-klein-9bBFLself-host-vastLoRA$1.00$1.00
bagelByteDanceself-host-vastLoRA + FT$2.00$2.00
step1x-editStepFunself-host-vastLoRA$2.00$2.00
flux-1-kontext-devBFLfalLoRA$2.50$2.50
qwen-image-editAlibabafalLoRA$4.00$4.00
flux-2-devBFLfalLoRA$8.00$8.00
hunyuan-image-3.0-instructTencentself-hostLoRA$75.00$75.0080B MoE — needs multi-H100

self-host on H100 · estimated cost for 8B × 1,000 steps

spot pricing
providergpukind$/hrtotal (0.28h)regionnotes
Vast.aiH100 80GBmarketplace$0.34$0.09globalMarket floor. Reliability varies by host.
Spheron / GMIH100 80GBmarketplace$0.47$0.13globalDecentralized GPU marketplace.
RunPodH100 SXM 80GBcloud$1.19$0.33global
SF ComputeH100 80GBcluster$1.96$0.54us-sfoVetted clusters, sell-back unused. B300s announced for fall 2026.
RunPodH100 PCIe 80GBcloud$1.99$0.55global
AWSH100 80GB (p5)cloud$2.50$0.69globalWas $6.88; repriced mid-2025. $1.90/hr with 3-yr SP.
CoreWeaveH100 80GB (8x HGX)cloud$2.83$0.79usPer-GPU price in 8x node; $49.24/hr for full node
LambdaH100 SXM 80GBcloud$2.99$0.83us+euReserved 1-yr: $1.89/hr
ModalH100 80GBserverless$3.95$1.10globalPreemptible base; non-preempt ~3.75×. Includes Modal Sandbox runtime.
Replicate (raw GPU)H100 80GB (8x)serverless$5.49$1.53usPer-GPU in 8x cluster; $43.92/hr full node.
how this works

Managed — provider runs the GPU + trainer code; you upload a dataset + config, get back LoRA weights. Best for low-volume, no-ops use.
Self-host — you rent the GPU and run ai-toolkit / kohya / diffusion-pipe. Best at high volume — Vast.ai spot H100 ≈ $0.34/hr beats fal's $4/1k Qwen-Edit trainer at > 10 LoRAs/month if you've automated the loop.