Predibase

core-platform

closed(proprietary post-training stack + turbo serving)

source

partial(LoRAX and Ludwig are adjacent OSS

OSS-components

LoRAX(open source multi-LoRA serving)+Ludwig(open source)

runs-on

Predibase-SaaS-or-customer-VPC

license

Proprietary(managed) with Apache-licensed LoRAX/Ludwig adjuncts

Predibase's founders created and still maintain the open-source LoRAX multi-LoRA serving stack and Ludwig, but the product scored here is the fine-tuning and reinforcement-fine-tuning platform, and that is proprietary managed SaaS running on Predibase's own cloud or inside a customer VPC. The open source sits at the serving and declarative-ML edges: an open periphery around a closed core, rather than an open core with paid extras on top. Because public repositories exist while the runtime itself does not ship, the platform counts as source-available rather than open source.

• https://pypi.org/pypi/predibase/json recorded 2026-08-09

  The JSON API's info object records license: null, classifiers: [], summary: null, and home_page: null for the predibase package (version 2025.10.3) on PyPI — no license or project metadata is declared for the published client library.

• https://files.pythonhosted.org/packages/f2/f0/6fb3406af8891fe08b8e5d3b325c39050f7616054bf0a73496277701356a/predibase-2025.10.3.tar.gz recorded 2026-08-09

  Extracted with Python's tarfile (106 members). README.md opens "## Predibase Python SDK"; PKG-INFO records "License: UNKNOWN"; no LICENSE file exists anywhere in the archive; setup.py registers only a console-script entry point (pbase=predibase.cli:main) and no license classifier. A case-insensitive grep of every decoded file for "vpc" found exactly one hit, in predibase/resources/datasets.py: ":param region: (str) Optional region to upload to. Only relevant for VPC customers with multiple dataplanes." This corrects a prior run's claim that the sdist contains no VPC reference.

• https://web.archive.org/web/20260107043258/https://docs.predibase.com/ recorded 2026-08-09

  The "Additional Capabilities" list states "SaaS or VPC deployment options: Deploy in our cloud or yours" and "UI and SDK: Deploy, fine-tune, and prompt via our UI or Python SDK" — access is through a hosted UI/SDK and a deployment-location choice, not a self-hostable code drop of the platform itself.

• https://github.com/predibase/lorax recorded 2026-08-09

  The repository title reads "Multi-LoRA inference server that scales to 1000s of fine-tuned LLMs"; the repo metadata records "license":{"spdxId":"Apache-2.0","name":"Apache License 2.0"} and "archived":false. The serving engine that plugs into Predibase is public, OSI-licensed, and actively maintained, in contrast to the managed fine-tuning/training platform itself, which ships no equivalent repository.

No hard per-platform usage figure -- jobs run or customers -- is disclosed. What can be reported is traction rather than volume: an enterprise post-training and agent platform now distributed under Rubrik, following an acquisition the press valued at $100M-$500M; founders out of Google and Uber; and a widely adopted open-source serving layer in LoRAX. That supports a modest level on reported enterprise traction rather than a measured download or customer count, which is why confidence is low.

• https://techcrunch.com/2025/06/25/rubrik-acquires-predibase-to-accelerate-adoption-of-ai-agents/ recorded 2026-08-09

  Reports the Rubrik acquisition of Predibase: "Terms of the deal were not disclosed, though CNBC reported that the deal was between $100 million and $500 million"; Predibase, founded in 2021 by Devvret Rishi (CEO), Piero Molino (chief science officer) and Travis Addair (CTO), "has raised more than $28 million in VC money" from Felicis, Greylock, and Sancus Ventures. No per-platform usage or customer count is given anywhere in the piece.

Strong managed method depth: supervised fine-tuning, continued pretraining and end-to-end reinforcement fine-tuning in the GRPO mold -- which Predibase pioneered for LLMs, learning from outcomes and feedback without heavy data preparation -- plus LoRA training and LoRAX multi-adapter serving that hosts hundreds of fine-tuned models per GPU, and turbo serving around twice as fast, all over open-source base models. That matches the reinforcement-fine-tuning services from OpenAI, Azure and Fireworks and adds best-in-class multi-adapter serving, but it is still a black-box managed platform with no exposed large-scale training internals, so it stays below Megatron-LM and the other open-source scale definers.

• https://web.archive.org/web/20260107043258/https://docs.predibase.com/ recorded 2026-08-09

  The Key Features section describes Serving as deploying models via production-ready private serverless endpoints through LoRAX, which serves hundreds of fine-tuned models on a single GPU, and Post-training as supervised fine-tuning, reinforcement fine-tuning, or continued pretraining. A features card elsewhere names Reinforcement Fine-Tuning specifically as Group Relative Policy Optimization (GRPO), and the docs' internal redirect map still routes a turbo_lora path to a TurboLoRA page, confirming it as a current feature name rather than a retired one.

• https://web.archive.org/web/20260118040413id_/https://predibase.com/ recorded 2026-08-09

  Searched the raw HTML including script payloads (not just visible text), case-insensitive substring counts: "fine-tun" appears 11 times and "TurboLoRA"/"reinforcement fine-tun" once each. The Organization JSON-LD lists knowsAbout entries for "Fine-tuning", "LoRA", "TurboLoRA", and "Reinforcement fine-tuning", and keywords "LLM, LoRA, inference, AI, fine-tuning, open-source models, AI serving". This directly contradicts a prior claim that this archived homepage has no fine-tuning mentions. The rendered page itself (title "Predibase: Govern Every Agent. Trust Every Action.") had already pivoted to agent-governance messaging by this date even though the schema block still lists the fine-tuning capabilities.