Top 7 LLM Fine-Tuning Platforms in 2026

Two well-known names in this category — Predibase and OpenPipe — were acquired in 2025 and no longer operate as independent products. Here's who's actually still standing, and worth using, today.

Fine-tuning has quietly become a cost play as much as a quality play: teams distill an expensive frontier-model prompt into a smaller, cheaper open-weight model once the task is well understood, rather than paying frontier-API rates forever. The tooling for doing that spans fully managed platforms that handle GPUs and serving for you, enterprise data-platform add-ons, and open-source frameworks you run yourself. This list covers the seven still-active options worth evaluating as of August 2026, and flags the acquisitions that quietly took two former category leaders off the table.

How we picked these

We prioritized platforms and frameworks that are currently available to sign up for or download independently — which meant dropping Predibase (acquired by Rubrik in mid-2025) and OpenPipe (acquired by CoreWeave in September 2025), both of which have been folded into their acquirers' broader platforms rather than continuing as standalone fine-tuning products. Among what remains, we ranked by adoption, breadth (does it also handle serving, at what cost), and whether pricing is published or requires a sales conversation.

Quick comparison

1. Together AI

Together AI offers supervised fine-tuning, DPO, and LoRA fine-tuning as part of a broader platform that also includes serverless inference and raw GPU rental — a genuine one-stop option for teams that would rather not stitch together separate vendors for training and serving. Fine-tuning is billed per million training tokens rather than per GPU-hour, with rates that scale by model size: small models run under $0.50 per million tokens, while full fine-tunes of 70–100B-parameter models run several dollars per million tokens (LoRA fine-tunes cost less than full fine-tunes at the same model size). Once trained, models can be served through Together's hosted endpoints or exported.

Best for: Teams that want fine-tuning and inference from the same managed vendor, minimizing the number of accounts and integrations in the stack.

Pros - Fine-tuning, hosted inference, and raw GPU rental all live under one account - Supports SFT, DPO, and LoRA — not just one fine-tuning method - Published per-token pricing that scales predictably with model size - Broad open-model catalog for both base models and serving

Cons - Per-token training pricing can be harder to estimate upfront than a flat per-GPU-hour rate, especially for iterative experimentation - LoRA adapter-serving flexibility is less developed than Fireworks AI's dedicated-deployment adapter model - Full fine-tunes of large (70B+) models get expensive quickly relative to LoRA-only specialists - Less focused than single-purpose fine-tuning tools — the platform's breadth can mean less depth in any one workflow

2. Fireworks AI

Fireworks AI built its fine-tuning product around LoRA from the start, and it shows in the pricing: fine-tuned models deploy at the same token price as the base model, with no serving surcharge — a real differentiator, since several competitors charge more to serve a fine-tuned model than its base version. Fireworks supports up to 100 LoRA adaptations running simultaneously on a single dedicated deployment at no extra cost, and offers two deployment modes — merging LoRA weights directly into the base model for lowest latency, or stacking multiple LoRA "addons" on one deployment to serve many fine-tuned variants from shared infrastructure. Training is priced per million training examples, with a Mixtral-class rate around $2 per million examples as a reference point.

Best for: Teams that need to serve many task-specific LoRA adapters cheaply without paying a fine-tuned-model serving premium.

Pros - No serving surcharge for fine-tuned models — they run at base-model token pricing - Up to 100 LoRA adapters can run on a single dedicated deployment - Two deployment modes (merged vs. addon) give real flexibility between latency and adapter density - Training API is straightforward and well-documented for LoRA-based SFT

Cons - Full (non-LoRA) fine-tuning is less emphasized than the LoRA-first workflow - Multi-adapter serving requires understanding the merge-vs-addon tradeoff, which adds a learning curve versus simpler single-model platforms - Smaller open-model catalog and community footprint than Hugging Face's ecosystem - Enterprise-scale support and SLAs are less established than Databricks' or Together AI's

3. Databricks Mosaic AI

Databricks Mosaic AI is the enterprise-lakehouse path into fine-tuning, built on the 2023 acquisition of MosaicML for roughly $1.3 billion. Rather than a standalone fine-tuning product, it's a feature inside the broader Databricks platform: fine-tuning, Foundation Model APIs, model serving, and vector search all bill through Databricks Units (DBUs) — a normalized software-usage metric layered on top of separately billed cloud compute from AWS, Azure, or GCP. Model training and fine-tuning are priced at roughly $0.65 per DBU, notably higher than the roughly $0.07 per DBU baseline rate and about ten times the model-serving DBU rate, reflecting the heavier compute fine-tuning jobs require.

Best for: Enterprises that already run their data pipelines on Databricks and want fine-tuning to live next to the data rather than in a separate vendor.

Pros - Fine-tuned models integrate directly with existing Databricks data pipelines, governance (Unity Catalog), and MLOps tooling - Backed by the MosaicML team's original training infrastructure and research - One consumption-based billing model (DBUs) spans fine-tuning, serving, and the rest of the Databricks platform - Strong fit for regulated industries already standardized on Databricks for compliance and data governance

Cons - DBU pricing is opaque to teams not already fluent in Databricks' billing model, and underlying cloud compute is billed separately, on top of the DBU charge - Fine-tuning DBU rates are meaningfully higher than serving rates, making iterative experimentation costlier than on token-metered platforms - Only makes sense as a fine-tuning platform if you're already invested in the broader Databricks ecosystem — it's not a good fit as a standalone tool - Less developer-friendly for quick, small-scale experiments than Hugging Face AutoTrain or the open-source frameworks on this list

4. Hugging Face AutoTrain

Hugging Face AutoTrain is the no-code entry point into fine-tuning: upload a dataset, pick a base model, set a few hyperparameters, and the platform runs the training job on managed H100 or L40S hardware. There's no subscription layered on top — you pay only for the compute minutes the run consumes, and simple fine-tunes can cost under $1 and finish in around 10 minutes. It covers more than LLM text generation, extending to text classification, named-entity recognition, semantic search, translation, and image classification, which makes it a reasonable default for teams without a dedicated ML infrastructure function.

Best for: Small teams, product managers, and anyone who wants the fastest and cheapest path to a first fine-tune without writing training code.

Pros - No subscription fee — pure pay-per-compute-minute pricing - Genuinely no-code: upload data, pick a model, set hyperparameters, run - Covers multiple task types beyond text generation (classification, NER, translation, image classification) - Deep integration with the wider Hugging Face model and dataset ecosystem

Cons - Limited configurability compared to Axolotl or Unsloth for teams that want to control training internals (learning-rate schedules, custom loss functions, exotic architectures) - Best suited to smaller, well-understood datasets — not built for frontier-scale training runs - Managed hardware selection is narrower than what you'd get provisioning your own GPUs on Modal or a neocloud - Less useful for teams that need production serving infrastructure baked in, since AutoTrain's focus stops at the training job itself

5. Modal

Modal isn't a fine-tuning product in the narrow sense — it's a general-purpose serverless compute platform that has become a popular substrate for teams that want to write and run their own fine-tuning pipelines without managing servers. Billing is fully consumption-based and per second: roughly $0.001036/second for an A100 40GB and $0.002778/second for an H100, with automatic scale-to-zero when nothing is running and $30 of free compute credit included monthly. Because there's no managed fine-tuning workflow, teams bring their own training code (commonly built on Hugging Face's Transformers or Axolotl) and let Modal handle provisioning, scaling, and per-second billing around it.

Best for: Engineering teams that want full control over a custom fine-tuning pipeline while avoiding the overhead of managing their own GPU infrastructure.

Pros - True per-second billing with automatic scale-to-zero — no idle GPU cost - $30/month in free compute credits lowers the barrier for small teams to start - Works with any training code or framework (Axolotl, Unsloth, raw Transformers) rather than locking you into a proprietary training API - Fast to get a custom pipeline running relative to standing up your own Kubernetes cluster

Cons - No built-in fine-tuning workflow — you write and maintain the training code yourself - Region and reliability multipliers (1.25x–3x over base per-second rates in some cases) can push real costs well above the headline rate - Less turnkey than Together AI, Fireworks, or Databricks for teams that want a managed, click-to-train experience - No native model-serving product comparable to a dedicated inference platform — you're responsible for deploying the fine-tuned result elsewhere

6. Axolotl

Axolotl is a free, open-source (Apache 2.0) fine-tuning framework originally built by the OpenAccess AI Collective, and it's become one of the most widely used tools for teams that want full control over their training runs without building infrastructure from scratch. Configuration is YAML-based, it supports LoRA and QLoRA alongside full fine-tuning across a wide range of model architectures, and it integrates with both Weights & Biases and the open-source MLflow for experiment tracking. Because it's a framework rather than a hosted service, there's no Axolotl-specific bill — you run it on whatever compute you provision, whether that's your own hardware, a neocloud GPU rental, or a serverless platform like Modal.

Best for: Teams that want maximum configurability over open-weight training runs and are comfortable managing their own compute.

Pros - Free and Apache 2.0 licensed, with no vendor lock-in - YAML-based configuration supports both quick starts and deep customization - Broad model architecture support and an active open-source community - Works with your choice of experiment tracking (Weights & Biases or MLflow) and your choice of compute provider

Cons - No hosted infrastructure — you're responsible for provisioning and managing the GPUs it runs on - No managed serving layer; once training finishes, deploying the result is a separate step - Steeper learning curve than no-code platforms like Hugging Face AutoTrain - Community-driven support model rather than a vendor SLA or dedicated enterprise support line

7. Unsloth

Unsloth is a Y Combinator-backed, open-source fine-tuning library built around one core promise: dramatically lower VRAM requirements and faster training than standard Hugging Face fine-tuning workflows, letting individual developers fine-tune or run reinforcement learning on free tiers of Google Colab or Kaggle, or locally with as little as 3GB of VRAM. The core library is Apache 2.0 licensed; an optional Unsloth Studio UI component is separately licensed under AGPL-3.0. It's built for accessibility as much as performance — the free-tier-Colab workflow is a deliberate design choice aimed at solo developers and small teams without dedicated GPU budgets.

Best for: Solo developers and small teams fine-tuning open-weight models on limited hardware, including free-tier cloud notebooks.

Pros - Runs meaningfully lower VRAM requirements than standard fine-tuning workflows, down to 3GB locally for some configurations - Free to use, with a genuinely usable free-tier Colab/Kaggle workflow, not just a marketing claim - Apache 2.0 core license keeps the primary library open for commercial use - Backed by Y Combinator, suggesting continued development investment

Cons - Optional Unsloth Studio UI is AGPL-3.0 licensed, which carries different (more restrictive for commercial redistribution) obligations than the Apache-licensed core - Optimized primarily for LoRA/QLoRA-style efficient fine-tuning rather than large-scale full fine-tuning - No hosted serving product — like Axolotl, deployment of the fine-tuned model is left to the user - As a smaller, newer company relative to Hugging Face or the larger managed platforms, longer-term support commitments are less established

How to choose

If you want fine-tuning and serving handled by one vendor with predictable per-token pricing, start with Together AI or, if LoRA-adapter density and zero serving surcharge matter most, Fireworks AI. If your organization already runs on Databricks, Mosaic AI keeps fine-tuning inside the same governance and billing boundary rather than adding a new vendor. For a first fine-tune with no ML infrastructure team behind it, Hugging Face AutoTrain is the fastest, cheapest on-ramp. Teams that want to own their training pipeline but not their servers should look at Modal; teams that want to own the whole stack, including the training code, should use Axolotl or Unsloth — Axolotl for breadth and configurability, Unsloth when VRAM and cost are the binding constraint.

Frequently Asked Questions

What happened to Predibase and OpenPipe?

Predibase was acquired by data-security company Rubrik in June 2025 for a reported $100–500 million, and its technology has been folded into Rubrik's agentic AI offerings rather than continuing as a standalone fine-tuning product. OpenPipe was acquired by CoreWeave in September 2025; its platform has migrated into CoreWeave and Weights & Biases, with its reinforcement-learning library continuing under CoreWeave's roadmap.

Is LoRA fine-tuning enough, or do I need a full fine-tune?

LoRA (and its quantized variant QLoRA) is sufficient for most task-specific adaptation and costs a fraction of a full fine-tune, which is why most of the platforms on this list default to it. Full fine-tuning is generally reserved for cases where you're changing a model's behavior more fundamentally than a low-rank adapter can capture.

Do I need a GPU to fine-tune with Unsloth or Axolotl?

Yes, but Unsloth in particular is designed to work within free-tier cloud GPU allocations (Google Colab, Kaggle) or as little as 3GB of local VRAM for smaller models, which removes the need to provision dedicated infrastructure for a first experiment.

How much does it actually cost to fine-tune a small model?

On managed platforms billed per token, small models can be fine-tuned for well under $1 per million training tokens; Hugging Face AutoTrain has advertised full training runs completing in around 10 minutes for under $1 in compute. Costs scale up sharply with model size and full (non-LoRA) fine-tuning.

Should I use a managed platform or an open-source framework?

Managed platforms (Together AI, Fireworks AI, Databricks Mosaic AI, Hugging Face AutoTrain) trade some cost and flexibility for speed and less operational overhead. Open-source frameworks (Axolotl, Unsloth) cost nothing themselves but require you to provision and manage the compute, plus handle serving separately once training finishes.

Editor's note — sources:

Pricing and product details for Together AI, Fireworks AI, Hugging Face AutoTrain, Modal, Axolotl, and Unsloth are drawn from each company's or project's own pricing and documentation pages. Databricks Mosaic AI DBU pricing is drawn from Databricks' published pricing documentation. The Predibase and OpenPipe acquisitions are confirmed via TechCrunch, Rubrik's and CoreWeave's own newsroom announcements, and CNBC's reporting on deal terms. All figures current as of August 2026; verify against each vendor's live pricing page before making a purchasing decision.