Multi-tenant SaaS, hundreds of customer adapters
Serverless multi-LoRA at base-model per-token pricing. LoRAX-backed, Turbo LoRA for speedup. Cheapest path to ship per-customer fine-tunes.
★ PlatformsDirect LLM · vendor-neutralProduction grade
FINE-TUNING · VENDORSTwelve fine-tuning platforms. Ranked by what each actually does well.
Managed APIs, serverless multi-LoRA, BYO GPU clusters, on-prem NIM. The 2026 vendor landscape with pricing, residency, audit, and the one we'd pick for each shape of project.
OpenAIAnthropicGoogle VertexAWS BedrockAzureTogetherPredibaseModal
Vendors compared
12 (managed + BYO)
Default managed
Together AI or Predibase serverless
Default RFT
OpenAI RFT (o4-mini)
On-prem default
NVIDIA NeMo Customizer
[SIDE BY SIDE]
The matrix.
Pricing snapshot is mid-2026. Numbers move quarterly; we re-validate every engagement. The qualitative columns (base models, methods, residency) are the durable parts of the comparison.
Twelve platforms, side by side.
Pricing is mid-2026 and moves quarterly; the qualitative columns (bases, methods, residency) are the durable parts.
| Vendor | Base models | Methods | Pricing | Deployment | Residency |
|---|---|---|---|---|---|
OpenAI fine-tuning + RFT | gpt-4.1, 4.1-mini, 4.1-nano, gpt-4o, o4-mini (RFT) | SFT + RFT (GRPO) | $25/1M tokens SFT (gpt-4.1); $100/hr RFT, $5k/job cap | Managed in-platform endpoints | US (default), EU via Azure |
Anthropic Claude (via Bedrock) | Claude 3 Haiku only (GA Nov 2024) | SFT | Bedrock fine-tune pricing + Provisioned Throughput | AWS Bedrock, requires Provisioned Throughput | US, EU regions |
Google Vertex AI | Gemini 2.5 Pro / Flash / Flash-Lite | SFT + preference tuning (DPO-style) | Per training token + 1.5x base inference for tuned | Vertex AI endpoints | US, EU, asia-* regions |
AWS Bedrock | Bedrock-supported models + Custom Model Import (any HF model) | Managed SFT + Custom Model Import | Per token managed; $0.0785/min/CMU for imports | Bedrock endpoints, Provisioned Throughput for tuned | All AWS regions |
Azure AI Foundry | GPT-4.1, 4.1-mini, o-series (RFT) | SFT + RFT (mirrors OpenAI) | Mirrors OpenAI; $100/hr RFT o4-mini | Standard, Global Standard, Provisioned Throughput | Azure global regions |
Databricks Mosaic AI | Llama 3, Mistral, DBRX | Full SFT, LoRA, DPO | Serverless H100 with InfiniBand, ~10x lower than proprietary per Databricks | Unity Catalog governance + Model Serving | Databricks regions |
Together AI Our default | Open models up to 100B | LoRA, full SFT, DPO | $0.48 / $1.20 (16B); $1.50 / $3.75 (17-69B); $2.90 / $7.25 (70-100B) per 1M tokens | Serverless multi-LoRA, dedicated endpoints | US (default), EU on request |
Predibase Multi-tenant SaaS | Llama, Mistral, Qwen, Gemma, Phi | LoRA, RFT, DPO, KTO | Serverless at base-model per-token; Turbo LoRA add-on | LoRAX multi-adapter serving, VPC | US + VPC anywhere |
NVIDIA NeMo Customizer On-prem default | Llama, Mistral, Nemotron, custom | LoRA, P-tuning, full SFT, DPO, GRPO | On-prem (compute is yours) | Kubernetes + NIM via NIM Operator 2.0 | Your data center |
Modal | Any (BYO training script) | Any (PyTorch, Unsloth, Axolotl) | A10G $1.10/hr, H100 ~$3.95/hr, B200 clusters available | Serverless Python, schedule 128 B200s in one line | US (default), EU on request |
Lambda Labs | Any (BYO training script) | Any | 1-Click Clusters: $4.49/GPU-hr, 1-week minimum, no egress | 16 to 512 H100s, InfiniBand 400 Gb/s | US |
HuggingFace TRL + AutoTrain | Any open base on the Hub | SFT, RM, DPO, GRPO (TRL v1.0) | Open source; AutoTrain managed pay-per-use | Inference Endpoints or self-hosted | Hub-hosted (US/EU) or self-hosted |
Highlighted rows are our default picks for the most common project shapes.
[OUR DEFAULT PICKS]
What we recommend by shape of project.
Reasoning fine-tune, math + code + tool use
$100/hr, $5k cap per job. Verifier in the loop. The fastest path to a measurable reasoning lift if your data fits the o-series.
On-prem, regulated industry, sovereign weights
Kubernetes-native, NIM Operator 2.0 for serving, full method coverage (LoRA, SFT, DPO, GRPO). The enterprise on-prem default.
70B+ fine-tune on a tight budget
QLoRA on a single 48GB GPU via Unsloth (2x faster, 70% less VRAM). Modal's per-second billing matches the iteration loop.
EU residency + Claude voice
Only path to fine-tuned Claude. EU region, Provisioned Throughput. Good when the use case specifically needs Claude.
Open ecosystem, full control
TRL v1.0 unifies SFT, RM, DPO, GRPO. Lambda 1-Click for the GPUs. The recipe we recommend when no vendor lock-in is acceptable.
[WHAT YOU GET]
What we leave at handoff.
1 audit
Vendor selection memo, signed
1 setup
Account, BAA/DPA, residency configured
1 run
First fine-tune, eval-gated, deployed
Exit
Weights exportable, no lock-in
[COMMON QUESTIONS]
What buyers ask before they sign.
- We need EU residency. Who's the right choice?
- Vertex AI europe-west and Mistral La Plateforme are the most mature with both regional residency and customer-managed keys. AWS Bedrock Frankfurt is solid for Claude 3 Haiku fine-tunes and Custom Model Import. Azure West Europe mirrors OpenAI. For self-hosted on EU soil, Modal and Lambda both offer EU regions on request.
- OpenAI RFT or self-hosted GRPO?
- OpenAI RFT is faster to start, the bill is capped ($5k per job), and it runs on closed o-series. Self-hosted GRPO requires a rollout server and verifier infra, can run on any open base, and stays in your VPC. Pick by data sensitivity and base-model preference. For most regulated industries, self-hosted wins.
- Why do you recommend Predibase or Together by default?
- Serverless multi-LoRA at base-model per-token pricing is the cheapest path for SaaS products with hundreds of customer adapters. Predibase ships Turbo LoRA (speculative decoding) and LoRAX (open source) for thousands of adapters per GPU. Together has the broadest open-model coverage and the cleanest DPO pricing.
- Anthropic fine-tuning — when is Claude 3 Haiku the right answer?
- Narrowly. Claude 3 Haiku is the only Claude model with fine-tuning, available only through Bedrock, text-only, 32k context. It's a good choice when you specifically want Claude's voice in a tuned model and Provisioned Throughput cost is acceptable. For most projects we tune an open base instead.
[RELATED FINE-TUNING TOPICS]
Worth a look next.
01 · FINE-TUNING
Methods
SFT, LoRA, QLoRA, DoRA, DPO, SimPO, ORPO, KTO, GRPO/RFT, distillation, model merging. Every named technique with when it earns the build.
Read more02 · FINE-TUNING
Data pipeline
Sourcing, PII redaction (Presidio), synthetic data (Distilabel, Nemotron), DEITA quality scoring, MinHash + SemDedup, labeling vendors, feedback loops.
Read more04 · FINE-TUNING
By data + compute scale
Under 1k examples to over 1M, single A10G to 128 B200. Indicative cost, recommended method, hardware tier.
Read more05 · FINE-TUNING
Custom model build
Continued pretraining, SFT, preference optimization (DPO, SimPO, ORPO), reasoning distillation (R1 lineage), model merging (TIES, DARE). The full build pipeline.
Read more06 · FINE-TUNING
Compliance
EU AI Act (Article 25 substantial-modification trap), GDPR, HIPAA, FedRAMP, Colorado AI Act, India DPDP, China GenAI Measures. Region-by-region for tuned LLMs.
Read moreFINE-TUNING · VENDORS · KENSINK LABS
Bring the use case. We will pick the vendor.
Independent benchmark, residency-aware, BAA/DPA negotiated, deployment hardened. Sized to the scope, scoped to the audit, signed at the artifact.