Fine Tuning

Tool	Category	Segment	Platform / Tool	Plan / License	Monthly Price USD	Pricing Model	Free Tier / OSS	Included Usage / Limits	Model / Modality Support	Tuning Methods	Dataset / Eval Workflow	Integrations / Frameworks	Deployment / Hosting	Security / Privacy	Team / Governance	Best Fit	Main Limits / Caveats
Vertex AI Gemini Supervised Tuning No tagline	Fine Tuning	Cloud provider tuning	Google Vertex AI	Google Cloud pay-as-you-go	Usage-based; exact regional model pricing must be checked in Vertex AI pricing	Google Cloud tuning and endpoint usage	Google Cloud free credits may apply to new accounts, not a durable tuning free tier	Vertex docs cover supervised fine-tuning data preparation for Gemini models	Gemini models on Vertex AI; enterprise region/model availability varies	Supervised fine-tuning	Cloud Storage datasets, validation, tuning jobs and Vertex model resources	Vertex AI, Google Cloud IAM, Model Garden, pipelines and enterprise data stack	Google Cloud regional managed service	Google Cloud IAM, VPC-SC/regional controls and enterprise compliance options	Cloud project/IAM/billing governance	Enterprise Gemini customization inside Google Cloud controls	More setup than Gemini Developer API; cost/region/model constraints need current GCP check
Unsloth OSS No tagline	Fine Tuning	Efficient fine-tuning library	Unsloth	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS library; paid Pro/Enterprise features may exist separately	✓	Local resources cite Unsloth as 2-5x faster and lower-memory LLM finetuning	Popular open LLM families such as Llama, Mistral, Qwen, Gemma and related models depending release	LoRA/QLoRA-style SFT, DPO and RL-style recipes depending examples/version	Ready-to-use notebooks/templates, chat templates and HF dataset workflows	Transformers, TRL, PEFT, bitsandbytes, Colab/Kaggle/local GPU environments	Local/notebook/cloud GPU self-hosting	Data stays in chosen notebook/compute environment unless APIs/HF uploads are used	No SaaS governance in OSS	Fast low-memory fine-tuning on limited GPUs or free notebook environments	Hardware/model compatibility changes quickly; production users should pin versions
TRL OSS No tagline	Fine Tuning	RLHF / SFT library	Hugging Face TRL	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS training library; compute/model storage separate	✓	TRL provides trainers for supervised fine-tuning and alignment workflows in the Transformers ecosystem	Transformer LLMs and VLM extensions through HF ecosystem	SFT, reward modeling, PPO, DPO, ORPO, GRPO and related alignment methods depending version	Dataset processing, trainers, eval hooks and model publishing through Hugging Face tooling	Transformers, PEFT, Accelerate, Datasets, Hub, Weights & Biases and experiment tooling	Local, notebook, cluster or cloud GPU training	Data stays local if self-hosted; hub upload visibility must be managed	No SaaS governance by default; HF org controls if using Hub	Researchers and ML engineers implementing alignment/fine-tuning recipes	Requires ML training expertise and GPU budget; not a no-code product
Axolotl OSS No tagline	Fine Tuning	Config-driven fine-tuning	Axolotl	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS YAML-config training framework	✓	Local resources describe Axolotl as an open-source framework for fine-tuning and evaluating LLMs	Open LLMs from HF ecosystem; multi-GPU and quantized setups depending config	SFT, LoRA, QLoRA, full fine-tuning, DPO and other recipes depending version	YAML configs for dataset preprocessing, training, inference and evaluation	Transformers, PEFT, Accelerate, DeepSpeed, bitsandbytes, W&B and HF Hub	Local/cloud/cluster GPU training	Data privacy depends on training environment and tracking integrations	No SaaS governance by default	Teams that want reproducible fine-tuning configs and model sharing	Config complexity can be high; exact model recipes need validation
LLaMA-Factory OSS No tagline	Fine Tuning	Fine-tuning framework	LLaMA-Factory	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS framework with GUI/CLI/API; compute separate	✓	Local resources list LLaMA-Factory as unified efficient fine-tuning for 100+ LLMs	Large catalog of open LLMs and multimodal models depending release	Pretraining, SFT, LoRA/QLoRA, DPO, PPO/KTO/ORPO and related alignment methods	Dataset templates, config-driven training, evaluation, export and chat UI workflows	Transformers, PEFT, TRL-like methods, bitsandbytes, DeepSpeed, Accelerate and HF Hub	Local, workstation, notebook or cluster GPU training	Data stays in chosen training environment	No SaaS governance by default	Practitioners who want broad model coverage and many tuning methods in one framework	Large feature surface; config/version management matters
torchtune OSS No tagline	Fine Tuning	Native PyTorch fine-tuning	torchtune	BSD-style / open source	$0 software; GPU/hosting costs separate	OSS PyTorch library	✓	Local resources list torchtune as a native PyTorch library for LLM fine-tuning	Open LLM recipes in PyTorch ecosystem; model coverage follows torchtune releases	SFT, LoRA/QLoRA and preference/alignment recipes depending release	Recipe-based configs, datasets, checkpoints and evaluation utilities	PyTorch, TorchAO, Hugging Face model formats and distributed PyTorch tooling	Local/cloud GPU training	Data stays in chosen training environment	No SaaS governance by default	PyTorch-first teams wanting readable, hackable fine-tuning recipes	Less turnkey than GUI frameworks; recipe support can lag newest models
LitGPT OSS No tagline	Fine Tuning	Training and deployment framework	LitGPT	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS framework; cloud compute separate	✓	Local resources describe LitGPT as pretraining, finetuning and deploying 20+ LLMs	Open LLM families supported by LitGPT recipes	Pretraining, SFT, LoRA/QLoRA and deployment/export workflows	Recipes for prepare-data, finetune, evaluate, quantize and deploy	PyTorch Lightning ecosystem, HF checkpoints, quantization and deployment tools	Local/cloud GPU training	Data privacy depends on training environment	No SaaS governance by default	Engineers wanting a compact end-to-end train-finetune-deploy framework	Model coverage and recipe maturity vary by release
H2O LLM Studio OSS No tagline	Fine Tuning	No-code fine-tuning UI	H2O LLM Studio	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS GUI/workflow; infrastructure separate	✓	Local resources list H2O LLM Studio as a no-code GUI/framework for fine-tuning LLMs	Open LLMs supported by the H2O LLM Studio environment	Supervised fine-tuning and experiment workflows	Dataset import, experiment setup, metrics, comparison and model export through UI	Hugging Face, Python ML stack and H2O ecosystem	Local/server GPU environment	Data stays in chosen deployment unless external integrations are used	No SaaS governance by default	Teams wanting GUI-driven LLM fine-tuning experiments	May be heavier to maintain than script-based recipes; model support needs checking
PEFT OSS No tagline	Fine Tuning	Parameter-efficient tuning	Hugging Face PEFT	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS library for parameter-efficient adapters	✓	PEFT supports parameter-efficient training methods that reduce trainable parameters and memory	Transformers models across text, vision and diffusion families depending adapter method	LoRA, QLoRA-style adapter workflows, prefix/prompt tuning and other PEFT methods	Adapter configuration, training with Transformers/TRL and adapter loading/merging	Transformers, Accelerate, TRL, bitsandbytes and Hugging Face Hub	Local, notebook, cluster or cloud GPU training	Data privacy depends on execution environment	No SaaS governance by default; HF Hub org controls if publishing	Fine-tuning open models when full-parameter training is too expensive	Needs compatible base model, quantization setup and careful adapter management
OpenAI Fine-Tuning Platform Wind-Down No tagline	Fine Tuning	Provider API	OpenAI	Existing users only / winding down	New user access closed; existing fine-tuning jobs temporarily available	Legacy fine-tuning/model optimization; current pricing page says platform is winding down	✕	OpenAI pricing page says fine-tuning platform is no longer accessible to new users; existing users can create jobs for coming months and models remain available until base deprecation	Selected OpenAI models only; exact availability follows deprecation timeline	Supervised fine-tuning, vision fine-tuning, direct preference optimization and reinforcement fine-tuning docs remain in model optimization section	Training files/jobs through OpenAI API for eligible existing orgs	OpenAI SDK/API and fine-tuning dashboard for eligible orgs	Hosted OpenAI API	OpenAI says fine-tuned models remain under customer control; API data controls apply	OpenAI organization/API key governance	Existing OpenAI fine-tuning customers maintaining already-planned tuned models	Not appropriate for new adopters because the platform is being wound down
Mistral Fine-Tuning API Legacy No tagline	Fine Tuning	Provider API	Mistral AI	Deprecated API feature	$4 minimum job fee; $2/month storage fee per model listed in deprecated docs	Legacy tuning job fees plus storage and inference costs	No durable free tier captured	Docs mark the feature deprecated and no longer actively supported; minimum fine-tuning fee and monthly storage fee are listed	Mistral text and vision fine-tuning docs are in deprecated resources	SFT for text and vision; classifier factory also listed in legacy docs	AI Studio / fine-tuning API dataset upload and validation	Mistral AI Studio/API and Mistral fine-tune codebase	Hosted Mistral API / AI Studio legacy path	Mistral API legal/privacy terms apply	Workspace/API key governance in Mistral console	Existing Mistral users with legacy fine-tune workflows	Deprecated status makes it risky for new production adoption
Fireworks Fine Tuning No tagline	Fine Tuning	Managed open-model fine-tuning	Fireworks AI	Usage-based	$0.50 / 1M training tokens and up for LoRA SFT; RFT by GPU hour	Per 1M training tokens for SFT/DPO; RFT by GPU hour; inference/hosting separate	New users receive free credits for platform usage, not a durable fine-tuning quota	Pricing page lists SFT/DPO by model size and says fine-tuned models serve at base-model price; RFT is billed by GPU hour	Open models up to very large parameter counts; text/vision fine-tuning options in docs	LoRA SFT, LoRA DPO, full-parameter SFT/DPO and reinforcement fine-tuning	Training jobs, image token accounting for VLMs, deployment of tuned models	Fireworks API, model library, LoRA deployments and open model workflows	Managed Fireworks training plus serverless/on-demand inference	Enterprise deployments and account controls available	Account/API key and enterprise controls	Teams tuning open models and deploying them on the same low-latency inference platform	Serving, deployment and training costs are separate; RFT cost depends on GPU time
OpenAI Reinforcement Fine-Tuning No tagline	Fine Tuning	Provider API	OpenAI	API feature for eligible models	$100/hour core training loop for o4-mini-2025-04-16; grader tokens billed separately	Wall-clock core training time plus model grader token usage	✕	Pricing page lists o4-mini RFT training at $100/hour and inference token rates for fine-tuned model usage	Reasoning-model optimization; current listed model is o4-mini-2025-04-16	Reinforcement fine-tuning with graders; supervised/preference routes differ	Grader design, evals and training jobs in OpenAI model optimization workflow	OpenAI SDK/API and eval/grader workflows	Hosted OpenAI API	API data controls and optional data-sharing discounts apply where enabled	OpenAI organization governance and access controls	Specialized reasoning tasks where reward/grader design is feasible	Expensive and specialized; model grader calls add separate inference cost
Gemini API Model Tuning No tagline	Fine Tuning	Provider API	Google Gemini API	API tuning feature	No fixed monthly fee captured; model/API billing and tuning availability apply	Tuning workflow for supported Gemini models; current costs should be checked in Google pricing pages	Gemini API free tier may apply to base API usage; tuning quotas/pricing need current project check	Gemini API docs describe model tuning flow through tunedModels and training examples	Supported Gemini models only; text tuning focus in Gemini API docs	Supervised model tuning	JSONL/example dataset preparation and tuned model creation through API/SDK	Google GenAI SDKs, AI Studio and REST	Gemini Developer API; Vertex AI is enterprise path	Google Developer API/AI Studio terms; data-use terms differ by tier	Google project/API key governance	Developers wanting first-party Gemini customization without self-hosting	Pricing and supported model list can change; verify current model and region before production
Together AI Fine-Tuning No tagline	Fine Tuning	Managed open-model fine-tuning	Together AI	Usage-based	$0.48 / 1M tokens and up for standard LoRA SFT by model size	Per 1M processed fine-tuning tokens; full fine-tuning and DPO priced higher; dedicated inference separate	No durable free fine-tuning tier captured	Pricing page lists standard fine-tuning per 1M tokens, including LoRA SFT up to 16B at $0.48 and larger/specialized model rows	Open-source models from Together catalog and selected custom/HF models; text, code, vision-language and specialized models	LoRA, full fine-tuning, DPO, function-calling, reasoning and VLM fine-tuning per docs	Dataset preparation, validation/eval dataset token accounting and dedicated endpoints after training	Together API/CLI, Hugging Face models, open-model ecosystem	Managed Together AI training and dedicated/serverless inference	Single-tenant/dedicated options and enterprise support available	Together account/team/billing controls	Managed open-model fine-tuning without operating GPUs	Fine-tuned inference may need dedicated endpoints; pricing varies by model size and method
AutoTrain No tagline	Fine Tuning	No-code / low-code training	Hugging Face AutoTrain	Open source / hosted HF service	$0 software; Spaces/compute/provider costs separate	OSS app plus Hugging Face compute or local infrastructure	✓	AutoTrain page lists LLM fine-tuning along with text, vision, tabular and sentence-transformer tasks	LLMs, VLMs, text classification, entity recognition, summarization, QA, translation, tabular and image tasks	LLM finetuning and broader ML task fine-tuning	No-code setup, dataset upload/configuration and model training jobs	Hugging Face Hub, Spaces, datasets, transformers and PEFT ecosystem	Local/self-hosted or Hugging Face-hosted compute	Data path depends on local vs HF-hosted execution and repo visibility	Hugging Face org, repo and billing governance	Beginners or teams wanting no-code fine-tuning experiments	Hosted cost depends on hardware; project maintenance/version should be checked before production
veRL OSS No tagline	Fine Tuning	Reinforcement learning framework	veRL	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS RL framework; compute separate	✓	Local resources list veRL as a flexible efficient reinforcement learning framework for LLMs	Open LLMs and agent/reasoning training workloads depending recipe	PPO, GRPO and RL-style post-training recipes depending release	Rollout, reward, trainer and distributed execution components	Ray, vLLM, PyTorch, Hugging Face and distributed training stacks	Self-hosted cluster/cloud GPU training	Data privacy depends on cluster/cloud setup	No SaaS governance by default	Teams experimenting with reasoning/agent RL post-training	Requires strong ML systems expertise and reward design
XTuner OSS No tagline	Fine Tuning	Fine-tuning toolkit	XTuner	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS toolkit	✓	Local resources describe XTuner as efficient, flexible and full-featured for fine-tuning large models	Open LLMs and multimodal models in InternLM/OpenMMLab ecosystem	SFT, LoRA/QLoRA and multimodal fine-tuning recipes depending version	Config-driven datasets, training configs and model conversion/export	Transformers, DeepSpeed, OpenMMLab and InternLM ecosystem	Local/cloud GPU training	Data privacy depends on training environment	No SaaS governance by default	Teams working with InternLM, LLaVA-style and Chinese/open-source model ecosystems	More ecosystem-specific than HF-native options
LLM Foundry OSS No tagline	Fine Tuning	Foundation model training code	Databricks LLM Foundry	Apache-2.0 / open source	$0 software; compute/platform costs separate	OSS training code; Databricks/Mosaic platform costs separate if used	✓	Local resources list LLM Foundry as LLM training code for Databricks foundation models	Open LLM pretraining/fine-tuning workloads in MosaicML/Databricks ecosystem	Pretraining, finetuning and evaluation recipes	Training configs, datasets, evaluation and model checkpoint workflows	PyTorch, Composer, MosaicML/Databricks stack and cloud GPUs	Self-hosted/cloud or Databricks/Mosaic platform	Data privacy depends on deployment environment	Governance through repo/self-hosting or Databricks controls if used	Teams wanting production-grade training code with Databricks/Mosaic lineage	More training-platform oriented than simple hobby LoRA workflows
mistral-finetune OSS No tagline	Fine Tuning	Vendor OSS fine-tuning codebase	mistral-finetune	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS codebase; compute separate	✓	Local resources list mistral-finetune as a lightweight codebase for memory-efficient Mistral model fine-tuning	Mistral open models supported by the repository	Memory-efficient supervised fine-tuning	Dataset formatting and training scripts for Mistral-family models	PyTorch, Transformers-style open model workflows	Local/cloud GPU training	Data stays in chosen training environment	No SaaS governance by default	Mistral open-weight model tuning where vendor-provided recipes are preferred	Focused on Mistral models; hosted Mistral API fine-tuning is deprecated separately
DeepSpeed OSS No tagline	Fine Tuning	Distributed training optimizer	DeepSpeed	MIT / open source	$0 software; GPU/hosting costs separate	OSS distributed training library; compute separate	✓	Local resources describe DeepSpeed as a deep learning optimization library for distributed training and inference	Large transformer models across PyTorch ecosystem	Enables full fine-tuning/pretraining at scale through ZeRO, offload and parallelism; not a fine-tuning UI by itself	Training optimization layer used inside recipes/frameworks	PyTorch, Hugging Face, Axolotl, LLaMA-Factory, Megatron and cluster training	Self-hosted/cloud GPU clusters	Data privacy depends on cluster/cloud setup	No SaaS governance by default	Scaling fine-tuning/pretraining beyond single-GPU memory	Requires distributed systems expertise; not a complete tuning product alone
Oumi OSS No tagline	Fine Tuning	End-to-end open model platform	Oumi	Apache-2.0 / open source	$0 software; compute/model costs separate	OSS end-to-end framework; compute separate	✓	Local resources describe Oumi as everything needed to build foundation models end-to-end	Open LLM and multimodal model workflows depending recipe	Training, fine-tuning, evaluation and deployment workflows	Config-driven datasets, training, evaluation and inference commands	Hugging Face, PyTorch, cloud/local training and open model ecosystem	Local/cloud GPU training and deployment	Data privacy depends on chosen environment	No SaaS governance by default	Teams wanting a unified OSS toolkit from data to evaluation/deployment	You still operate compute and validate recipes for each model
Megatron-LM OSS No tagline	Fine Tuning	Large-scale training framework	Megatron-LM	Open source	$0 software; GPU/hosting costs separate	OSS training framework; compute separate	✓	Local resources list Megatron-LM as ongoing research training transformer models at scale	Very large transformer models on NVIDIA GPU clusters	Pretraining and full fine-tuning style large-scale training; adapter workflows require integration	Distributed data/model/tensor/pipeline parallel training recipes	NVIDIA GPU stack, PyTorch, Transformer Engine and cluster schedulers	Self-hosted/cloud NVIDIA GPU clusters	Data privacy depends on cluster/cloud setup	No SaaS governance by default	Frontier-scale or research-scale training teams	Heavy infrastructure requirement; overkill for ordinary LoRA tuning
Transformer Lab OSS No tagline	Fine Tuning	Desktop LLM engineering app	Transformer Lab	Open source	$0 software; local/cloud compute separate	OSS desktop/app environment	✓	Local resources describe Transformer Lab as an open-source app for advanced LLM engineering: interact, train, fine-tune and evaluate on your computer	Local/open LLMs supported by app integrations	Fine-tuning and evaluation workflows for local LLM experimentation	GUI workflows for model interaction, training/fine-tuning and evaluation	Local LLM ecosystem and common model formats depending app version	Local workstation or connected compute	Can keep data local when models and training run locally	No SaaS governance by default	Individuals and small teams wanting a local UI for LLM fine-tuning experiments	Local hardware limits model size and training speed
OpenRLHF OSS No tagline	Fine Tuning	RLHF framework	OpenRLHF	Apache-2.0 / open source	$0 software; GPU/hosting costs separate	OSS RLHF framework; compute separate	✓	Local resources list OpenRLHF as scalable high-performance RLHF supporting large models, LoRA and preference methods	Open LLMs including large and MoE models depending hardware/config	SFT, reward modeling, PPO, DPO, KTO, iterative methods and LoRA/full tuning depending version	Distributed training scripts, preference datasets and reward model workflows	Ray, DeepSpeed, vLLM, Hugging Face and distributed GPU clusters	Self-hosted cluster/cloud GPU training	Data privacy depends on cluster/cloud setup	No SaaS governance by default	Research and production teams running RLHF/alignment at scale	Cluster complexity and GPU budget are high
NVIDIA NeMo OSS No tagline	Fine Tuning	Enterprise training framework	NVIDIA NeMo	Apache-2.0 / open source	$0 software; NVIDIA/cloud costs separate	OSS framework with enterprise NVIDIA ecosystem options	✓	Local resources include NeMo as a generative AI framework for researchers and PyTorch developers across LLMs, multimodal, ASR and TTS	LLMs, multimodal models, ASR, TTS and CV workflows	SFT, PEFT, RLHF/alignment and large-scale training workflows depending NeMo component/version	Data curation, training, evaluation and deployment integration across NVIDIA stack	PyTorch Lightning style stack, Megatron, NVIDIA NIM/NeMo services and GPU clusters	Self-hosted/cloud NVIDIA GPU infrastructure	Enterprise controls depend on NVIDIA/cloud deployment	NVIDIA enterprise and cluster governance options	Organizations standardizing on NVIDIA AI stack for tuning and deployment	Complex stack; best fit when NVIDIA infrastructure is already a constraint or preference