Best Fine-Tuning Platforms in 2026

Managed fine-tuning with real weight ownership: per-token pricing, the broadest open-weight catalog, and checkpoints you can download and leave with.

Together wins because it sits exactly where most teams need to be after the OpenAI wind-down: a managed API with none of the lock-in that made managed tuning risky. The catalog spans 270M to 480B parameters, from Gemma 3 270M up to Qwen3 Coder 480B, across Llama 3.x and 4.x (Scout and Maverick), Qwen 3 through 3.6, Gemma 3 and 4, DeepSeek, Mistral and Mixtral, Kimi K2, GLM 4.6 through 5.1, GPT-OSS, and NVIDIA Nemotron, and you can bring a model from the Hugging Face Hub that is not in the catalog. LoRA is the default training mode, full fine-tuning is available on a subset of the catalog, and both SFT and DPO are supported, with training context up to 262,144 tokens on select models. The decisive feature is the exit: final and intermediate checkpoints are downloadable, merged or as adapters for LoRA jobs, so you can serve on a Together dedicated endpoint or take the weights to your own vLLM cluster. Pricing is legible (per million training tokens, $4.00 job minimum), and entry costs are low enough to run honest experiments. The honest limits: no managed reinforcement fine-tuning, and the largest MoE families carry separate, much higher pricing.

Fine-tuning fused to a production inference platform: SFT, DPO, and managed reinforcement fine-tuning that deploys where it will be served.

Fireworks is the pick when the tuned model's destination is production traffic and you want training and serving on one platform. Supervised and preference tuning bill per token at rates close to Together's at the small end (LoRA SFT $0.50 per million tokens up to 16B), and it is the only managed platform in this roundup with reinforcement fine-tuning as a first-class product: you supply a grader function that scores outputs, billing runs per GPU-hour at on-demand deployment rates, and grpo, dapo, and gspo-token loss functions are built in. The pricing page's standout commitment is that fine-tuned models serve at the same price as base models, and the docs support multi-LoRA deployments plus importing LoRA adapters trained elsewhere. The constraints are real, though. The docs state that on-demand dedicated deployments are the only supported method for serving fine-tuned models, which means the serving floor is GPU rental: $7.00 per H100 or H200 hour, $10.00 per B200 hour. LoRA rank caps at 32. And weight download is not addressed in the fine-tuning docs, so if exit portability is a hard requirement, get it in writing before committing the training budget.

A frontier-lab training stack for rent: four primitives, your own loops in Python, their GPUs, LoRA only, weights downloadable.

Tinker is the most interesting product in this category and the right choice when the training loop itself is your edge. Instead of a job API, it exposes four primitives (forward_backward, optim_step, sample, and save_state), and you write the actual training loop in Python on your laptop while Thinking Machines schedules it across their GPUs and owns the infrastructure reliability. That makes supervised learning, RL, DPO, RLHF, and distillation all expressible without owning a cluster. The catalog is curated and rotates: at verification it ran from the 4B Qwen3.5-4B up to the 550B-parameter Nemotron-3-Ultra MoE, including Kimi-K2.6, DeepSeek-V3.1, Qwen3.5-397B-A17B, and GPT-OSS, several with vision input. Any checkpoint you save is downloadable through the API for serving anywhere. It went generally available on December 12, 2025 with an OpenAI-compatible sampling interface. The trade-offs keep it at three for a general audience: it is LoRA-only by design (the lab's own LoRA Without Regret research argues LoRA matches full fine-tuning for supervised runs on small-to-medium datasets, and for RL even at small ranks), the catalog rotates on the vendor's schedule (older Llama, Qwen, DeepSeek, and Kimi variants retired June 12, 2026, and no Llama model was listed at verification), and you are writing real training code, which is the point, but not everyone's job.

The open-source control answer: the widest method coverage in this roundup, one YAML per pipeline, on whatever GPUs you can get.

Axolotl is what you run when the platform question becomes an infrastructure question. It is Apache 2.0, at v0.17.0 as of June 3, 2026, with 12.2k GitHub stars, and its method coverage is the broadest here: full fine-tuning, LoRA, QLoRA, quantization-aware training (int8, int4, FP8, NVFP4, MXFP4), preference methods (DPO, IPO, KTO, ORPO), reinforcement learning (GRPO and GDPO), and reward modeling for outcome and process reward models. A single YAML config drives preprocessing, training, evaluation, quantization, and inference, and it scales from one GPU to multi-node via FSDP1 and FSDP2, DeepSpeed, Torchrun, and Ray, with Flash Attention 2, 3, and 4, Liger kernels, and sequence parallelism. Model support runs from GPT-OSS, Llama, Mistral, and Mixtral through vision-language models (Qwen2-VL, Pixtral, LLaVA, InternVL 3.5, GLM-4.6V) and audio models such as Voxtral. The cost is honest and structural: there is no platform fee because there is no platform. You bring GPUs, you debug OOMs and dataset schemas, and you own the failure modes. For regulated data, air-gapped environments, or teams training continuously enough that per-token markups compound, that trade is correct.

The middle path: framework-level control on serverless GPUs, billed per second, with zero idle cost and no cluster to own.

Modal is not a fine-tuning product, and that is exactly its value: it removes the worst part of the self-hosted path (owning GPU infrastructure) while keeping all of the control. GPUs bill per second with no idle charge: an H100 at $0.001097 per second (about $3.95 per hour), an A100 80GB at $0.000694 per second (about $2.50 per hour), a B200 at $0.001736 per second (about $6.25 per hour). Modal's maintained examples cover the whole arc: LoRA fine-tuning with Unsloth (the published example tunes Qwen3-32B on L40S GPUs with model, dataset, and checkpoint storage in Modal Volumes), GRPO reinforcement training with both TRL and verl, and an OpenAI-compatible vLLM service for serving the result. The Starter plan is $0 per month with $30 of free monthly credits, which is genuinely enough to fine-tune a small model before you pay anything, and volumes at $0.09 per GiB per month hold your checkpoints. It ranks fifth not because it is weak but because it is a different layer: you still write and maintain training code, there is no managed job API, no evaluation UI, and no tuning-specific support. For teams with the engineering habit but not the hardware, it is often the best per-dollar answer on this page.