Fine-Tuning as a Service
URL: https://vintagedata.org/blog/posts/fine-tuning-as-service
Author: Pierre-Carl Langlais & Yannick Detrois (Vintage Data)
Published: March 30, 2026
Summary
An evaluation of managed fine-tuning platforms (Tinker, Nebius Token Factory, Together AI, Fireworks, Prime Intellect) for demanding post-training workflows — specifically iterative synthetic data generation with specialized generator models. The benchmark is agentic SFT+RL, not simple instruction-tuning.
Key Claims
- The gap outside frontier labs: Most corporate/research teams attempting fine-tuning either give up or fall back to prompting — GPU cluster management and silent failures in long runs are the blockers.
- New managed platforms: Nebius Token Factory, Fireworks, Together AI, Prime Intellect now offer integrated SFT+RL infrastructure. The ecosystem is becoming modular — Cursor uses Fireworks for RL training of Composer-2.
- SYNTH methodology: Vintage Data’s fully synthetic generalist training environment trained SOTA small models without organic data. Moving to agentic use cases required qualitative step up in fine-tuning capability.
- Agentic traces are harder: Up to 30% of tool calls in early agentic training data had faulty JSON — quality issues that degrade model behavior.
- Why fine-tuning beats prompting at scale: Better performance on smaller models via specialization; accuracy on formal rules (math, structured output); style customization; built-in I/O structure; effective token use (no prompt needed).
- Specialization as moat: Fine-tuned specialist models outperform general models on domain tasks; bundled specialist generators are now how frontier labs create their own training data.
- Fireworks note: Used by Cursor for RL; appears in both this article and Fireworks Frontier RL as serious infrastructure player.
Evaluation Dimensions
Platforms compared on: general service offering (beyond strict fine-tuning), infrastructure/UI quality, cost-to-performance, model availability, and deployment of fine-tuned models.