What are the best platforms for LoRA fine-tuning on dedicated GPUs with managed infrastructure?

As of November 8, 2026: Together AI Post-Training when you want to write your own SFT or RL loop on managed B200s ($11 per GPU-hour, $9 promotional until December 14, 2026); Baseten Training Jobs when you want to bring your own Axolotl or TRL container on H100/H200 and deploy any checkpoint as an endpoint with one command; Fireworks' Dedicated Training API when the adapter will be served on Fireworks (H100 $8, B200 $13 per GPU-hour); Modal for per-second, scale-to-zero jobs; and Lambda or RunPod when you only need raw dedicated GPUs at the lowest hourly rate (RunPod H100 SXM $3.99, Lambda H100 $4.59 from November 16, 2026).

Should I pay per token or per GPU-hour for LoRA fine-tuning?

Pay per token (for example Together Fine-Tuning at $1.05 per million training tokens for Qwen3.8 27B, $4 minimum) for small, standard supervised jobs on supported models: you never touch a GPU. Pay per GPU-hour when you need an unsupported base model, a custom loop such as RL or DPO variants, long sequence lengths, or many runs a month; dedicated GPUs become cheaper once utilization is high and you control the recipe.

What GPU do I need for LoRA on a 70B model?

With QLoRA (4-bit base weights) a 70B model fits on a single 80 GB H100 or A100 for modest sequence lengths; plain 16-bit LoRA on 70B needs roughly two to four 80 GB GPUs or one 141 GB H200 or 180 GB B200 depending on context length and batch size. Models of 8B to 32B train comfortably on one 48 to 80 GB GPU.

Can I deploy the LoRA adapter on the same platform?

Yes on Baseten (baseten train checkpoint deploy turns a synced checkpoint into an endpoint), Fireworks (trained models are served on on-demand dedicated deployments) and Together (dedicated endpoints and serverless LoRA for supported bases). On Modal you deploy your own serving code. On Lambda and RunPod you run your own inference server such as vLLM.

Quick Answer

Best Platforms for LoRA Fine-Tuning on Dedicated GPUs

Published: November 8, 2026

The short answer

Managed dedicated-GPU training means you bring the recipe and the platform brings provisioned GPUs, checkpoints and a path to production. As of November 8, 2026:

Together AI Post-Training — best for custom SFT and RL loops on managed B200s.
Baseten Training Jobs — best bring-your-own-container training with one-command deployment.
Fireworks Dedicated Training — best when you will serve the adapter on Fireworks.
Modal — best for bursty jobs billed per second with no idle cost.
Lambda and RunPod — best raw dedicated GPUs at the lowest hourly price.

If your job is a standard supervised fine-tune on a supported model and you would rather pay per token, see QLoRA platforms with token billing instead.

The comparison

Platform	What “managed” covers	GPUs	Price (list, Nov 8, 2026)	Serving the result
Together AI Post-Training	You write the training loop (SFT or RL); Together runs the infrastructure	B200	$11 per GPU-hour; $9 promo until Dec 14, 2026; billed per minute per active resource	Together dedicated endpoints
Together GPU Clusters	Provisioned clusters, you run the stack	H100, H200, B200, B300	H100 $3.99 on-demand, $1.99 preemptible, $3.19 reserved 181+ days · B200 $8.19	Your own server
Baseten Training Jobs	Your container (Axolotl, TRL, VeRL, MS-Swift) on managed GPUs; checkpoints synced automatically	H100, H200; single or multi-node	Not shown on Baseten’s public pricing page when checked	`baseten train checkpoint deploy`
Fireworks Dedicated Training API	Managed SFT, DPO, ORPO or your own loop	H100, H200, B200, B300	H100 $8 · H200 $8 · B200 $13 · B300 $15 per GPU-hour	On-demand dedicated deployment (required for trained models)
Modal	Serverless containers, per-second billing, scale to zero	Wide range	Starter $0 + compute ($30 credits a month, 10 GPU concurrency); Team $250 + compute	Your own serving code on Modal
Lambda Instances / 1-Click Clusters	Dedicated instances or 16–2,000+ GPU clusters	A100, H100, B200	H100 $4.59 · B200 $7.69 · A100 80GB $3.19 per GPU-hour (from Nov 16, 2026) · B200 clusters $9.86 → $8.87	Your own server
RunPod Pods	Dedicated GPU pods, per-second billing	L40S → B300	L40S $1.09 · A100 80GB $1.79 · H100 SXM $3.99 · H200 $5.29 · B200 $7.99 · B300 $8.99 per hour	RunPod Serverless or your own

Prices from each vendor’s pricing page or documentation on November 8, 2026.

1. Together AI Post-Training — custom loops, no infrastructure

Together launched Post-Training on October 19, 2026 for building custom models with SFT or reinforcement learning: you define the training loop, Together operates the GPUs and optimizes the systems underneath. It bills per minute for each active model resource at $11 per B200 GPU-hour, cut to $9 until December 14, 2026. If you do not need a custom loop, Together’s token-billed Fine-Tuning (LoRA or full) is simpler: $1.05 per million tokens for Qwen3.8 27B, $2.03 for Llama 3.3 70B and $2.50 for gpt-oss-120B, with per-job minimums from $4.

Pick it if you want RL or a bespoke recipe on large open models without running a cluster.

2. Baseten Training Jobs — your container, managed GPUs

Baseten runs your own container and training code (Axolotl configs, TRL scripts for SFT, DPO and GRPO, VeRL for RL, MS-Swift for long context) on H100 or H200 GPUs, single or multi-node. Checkpoints sync to Baseten storage while the job runs, and baseten train checkpoint deploy turns any of them into a production endpoint, so there is no download-and-reupload step between training and inference.

Pick it if you already have a working recipe and the shortest path from a checkpoint to a serving endpoint matters most.

3. Fireworks Dedicated Training — train where you serve

Fireworks uses one GPU rate card for serving and training: the Dedicated Training API bills per GPU-hour at $8 for H100 or H200, $13 for B200 and $15 for B300; region-restricted deployments cost 1.5×. Managed training covers SFT, DPO and ORPO, and Fireworks’ docs state that an on-demand (dedicated) deployment is the only supported way to serve trained models, so budget for that too (idle deployments scale to zero after an hour by default).

Pick it if Fireworks will host the model in production.

Modal bills GPUs and CPUs per second and scales containers to zero, so a two-hour LoRA run costs two hours. The free Starter plan includes $30 of compute a month and 10 concurrent GPUs; Team costs $250 a month plus compute with $100 of credits and 50 concurrent GPUs. You write the training script in Python and Modal provisions the hardware.

Pick it if you run many short experiments and hate paying for idle machines.

5. Lambda and RunPod — cheapest dedicated GPUs

When “managed” only needs to mean “someone else racks the GPU”, hourly rentals win on price. RunPod lists an H100 SXM at $3.99/hour, H100 PCIe at $2.89, H200 at $5.29 and B200 at $7.99, billed per second (prices updated November 4, 2026). Lambda instances cost $4.59 per H100 GPU-hour and $7.69 per B200 from November 16, 2026, and 1-Click Clusters run 16 to 2,000+ GPUs for two weeks to a year from $9.86 per B200 GPU-hour. Together’s preemptible H100s at $1.99 are cheapest if your job checkpoints and can restart.

Pick it if you are comfortable running Axolotl, Unsloth or TRL yourself.

How to choose

Custom RL or SFT loop, no ops: Together Post-Training.
Existing container, fast path to an endpoint: Baseten.
Serving on Fireworks: Fireworks Dedicated Training.
Many short runs: Modal.
Lowest hourly price, you run the stack: RunPod, Lambda or Together preemptible.

For hosting the finished adapters, see best inference providers for LoRA fine-tuned models.

Last verified: November 8, 2026.