Which platforms offer QLoRA or LoRA fine-tuning with token-level billing in 2026?

As of October 2026 the three managed platforms that bill fine-tuning per token are Thinking Machines' Tinker (separate prefill, sample and train meters per million tokens, for example $4.10 per million train tokens on Qwen3.8-27B), Together AI (per million training tokens, for example $1.05 for Qwen3.8 27B SFT, with a minimum charge per job) and Fireworks AI (per training token for Managed Training and Serverless Training). All three train LoRA adapters. QLoRA itself — training adapters on a 4-bit quantized base — is a technique you run on rented GPUs with open-source trainers, billed per GPU hour rather than per token.

Can I still fine-tune GPT models on OpenAI?

Only if you are an existing customer. OpenAI is winding down self-serve fine-tuning: since May 7, 2026 organizations that never ran fine-tuning cannot start, since July 2, 2026 organizations without fine-tuned-model inference in the past 60 days cannot create jobs, and from January 6, 2027 no new fine-tuning jobs can be created. Fine-tuned models keep serving until their base models are deprecated.

What is the cheapest way to fine-tune an open model in 2026?

For small models, Together AI's per-token SFT rates start at $0.34 per million tokens (Llama 3.1 8B, Qwen3.5 up to 9B) with a $4 minimum per job, and Fireworks is running free LoRA SFT on Muse Glimmer 30B for jobs started before November 1, 2026. For large mixture-of-experts models, Tinker bills by active parameters and is running 50% launch discounts, for example $1.28 per million train tokens on Nemotron 3 Super.

Is per-token or per-GPU-hour billing better for fine-tuning?

Per-token billing is better for small and medium runs, experiments and RL loops with idle time, because you pay nothing between steps. Per-GPU-hour billing wins on long, steady jobs where you keep the GPUs busy, and is the only way to run QLoRA on a model or recipe the managed platforms do not support.

Quick Answer

Best QLoRA Fine-Tuning Platforms with Token Billing 2026

Published: October 15, 2026

The short answer

For LoRA fine-tuning billed by the token, use Tinker, Together AI or Fireworks AI; for true QLoRA, rent GPUs. Tinker (Thinking Machines) is the most flexible — you write the training loop, it runs it and bills prefill, sampling and training tokens separately. Together AI is the simplest managed SFT/DPO with a public per-model price list. Fireworks fits teams that will also serve the result there. OpenAI’s self-serve fine-tuning is closed to new customers and shuts to everyone on January 6, 2027. Prices are list USD as of October 15, 2026.

The comparison

Platform	Billing unit	Example train price (per 1M tokens)	Methods	Best for
Tinker	Prefill, sample and train tokens metered separately	Qwen3.8-27B $4.10 · Nemotron 3 Super $1.28 (50% off) · GLM-5.3 $14.58	LoRA with your own loop: SFT, RL, distillation	Custom recipes, RL, large MoE models
Together AI	Training tokens (dataset × epochs), minimum per job	Llama 3.1 8B $0.34 · Qwen3.8 27B $1.05 · gpt-oss-120B $2.50 · Kimi K2.6 $15	SFT and DPO; LoRA and full	Managed SFT/DPO with fixed prices
Fireworks AI	Training tokens (Managed and Serverless Training)	See pricing docs; Muse Glimmer 30B LoRA SFT free until Oct 31	SFT, DPO, ORPO, RFT, custom loops	Train and serve in one place
Self-hosted QLoRA	GPU hours	Your GPU rate	QLoRA on 4-bit bases, any recipe	Unsupported models, steady long jobs
OpenAI	Training tokens or hours	Closed to new orgs	SFT, RFT	Existing customers only, until Jan 6, 2027

Platform notes

Tinker. The pricing page lists three meters per model — prefill, sample and train — plus an 80% discount on cached prefill tokens and checkpoint storage at $0.10 per GB per month. MoE models are priced by active parameters, which makes very large models cheap to adapt: Nemotron 3 Ultra (550B total, 55B active) trains at $5.48 per million tokens during the current 50% promotion. Long-context variants cost more per train token (Qwen3.8-27B at 256K context is $7.46). Check the deprecation list before committing: Nemotron 3 Nano is retiring from Tinker on October 23, 2026.

Together AI. You pay for training dataset tokens times epochs, plus any evaluation tokens, with a minimum charge per model ($4 for small models, $60 for Kimi K2.6 and GLM-5.2). DPO costs about 2.5x SFT. It is the most predictable option: count your tokens, multiply, done. Serving the adapter afterwards means a dedicated endpoint — see inference providers for LoRA models.

Fireworks AI. Fireworks bills training per token for Managed Training and offers Serverless Training — a shared trainer pool for LoRA SFT, DPO, ORPO, RL and custom loops with no provisioning. Its documentation states trained models deploy only to dedicated (on-demand) deployments, so budget GPU hours for inference. Fireworks publishes a calculator comparing its dedicated training with Tinker’s per-token pricing.

Self-hosted QLoRA. If your base model or recipe is not on a managed list — or you specifically want 4-bit QLoRA to fit a large model on fewer GPUs — run an open-source trainer on rented GPUs. You pay per GPU hour whether the GPU is busy or not, so it pays off only on long, steady jobs.

How to choose

Standard SFT on a popular open model: Together AI — fixed per-token prices, small minimums.
RL, distillation or a custom loss: Tinker or Fireworks Serverless Training.
You will serve the adapter at scale: train where you deploy (Fireworks or Together) to skip export work.
Huge MoE model on a budget: Tinker, billed by active parameters.
Unsupported model or 4-bit QLoRA: self-host on GPUs.

Before you fine-tune, check whether a current frontier model with a good prompt is cheaper — see current API prices. This page expands the fine-tuning angle of our LLM inference provider ranking; the training process itself is in how to fine-tune LLMs.

Last verified: October 15, 2026. Prices are list USD from vendor pricing pages and may carry limited-time discounts.

The short answer

The comparison

Platform notes

How to choose

Sources