OpenTinker

license

Apache-2.0(OSI)

source

public(open-tinker/OpenTinker)

core features withheld

no — no ee, enterprise, commercial or proprietary path in the tree at commit f87fe25

training-backend

open(verl, Apache-2.0, included as a git submodule from volcengine/verl)

OpenTinker is Apache-2.0, and the client, scheduler and server all build from this repository. Its training backend, verl, is Apache-2.0 too, and the authors are university researchers who describe no paid edition or hosted tier.

• https://arxiv.org/pdf/2601.07376 recorded 2026-10-19

  Both authors, Siqi Zhu and Jiaxuan You, are listed at the "University of Illinois Urbana-Champaign". The paper presents OpenTinker as "an open infrastructure for multi-LoRA SFT and RL training of LLM agents" and says commercial systems "such as Tinker" keep their internal abstractions private.

• https://raw.githubusercontent.com/open-tinker/OpenTinker/main/.gitmodules recorded 2026-10-19

  One submodule, "verl", at path verl with url https://github.com/volcengine/verl.git.

• https://raw.githubusercontent.com/open-tinker/OpenTinker/main/LICENSE recorded 2026-10-19

  LICENSE body, read in full: the stock Apache License 2.0 text with its unfilled appendix boilerplate ("Copyright [yyyy] [name of copyright owner]") and no added term or condition.

• https://raw.githubusercontent.com/open-tinker/OpenTinker/main/README.md recorded 2026-10-19

  README installs the product from a clone ("git clone --recurse-submodules", "pip install -e .", then verl from its submodule) and describes Client Setup and Server Setup, with a built-in authentication system for the scheduler API. Searched the body for enterprise, pricing, paid and commercial: none.

• https://raw.githubusercontent.com/volcengine/verl/main/LICENSE recorded 2026-10-19

  LICENSE body of volcengine/verl: the Apache License, Version 2.0.

• https://ungh.cc/repos/open-tinker/OpenTinker/files/main recorded 2026-10-19

  The 190-file tree of open-tinker/OpenTinker at sha f87fe25fc483e7f1d7596182fc55115cbc992724: one root LICENSE, and no ee/, enterprise/, commercial/, proprietary/ or premium/ path anywhere in it. Path absence only, paired with the authorship statement cited beside it.

OpenTinker has no registry package: the opentinker name in its pyproject.toml is not published on PyPI, and the README installs it from a clone. GitHub stars are the only signal left, and a star records interest rather than a use.

• https://ungh.cc/repos/open-tinker/OpenTinker recorded 2026-10-19

  Repository metadata for open-tinker/OpenTinker: "stars":678, "forks":67, created 2025-12-20, pushedAt 2026-03-21.

OpenTinker trains agents with SFT and reinforcement learning in single- and multi-turn environments, letting many LoRA adapters share one base model with their training state kept apart, the same agent-RL ground OpenPipe's ART covers. Its documented examples stop at 3B-parameter models, so none of the large-scale parallel training of verl or Megatron-LM is shown.

• https://arxiv.org/abs/2601.07376 recorded 2026-10-19

  The abstract describes "multi-LoRA SFT/RL training in which many users, tasks, or agents can share a base model while keeping adapter updates, checkpoints, and rollout snapshots isolated", and "the training scheduler that isolates adapter-local gradients and optimizer state", validated on "single-turn, multi-turn, LoRA, and multi-agent agentic training".

• https://raw.githubusercontent.com/open-tinker/OpenTinker/main/docs/alfworld_multiturn.md recorded 2026-10-19

  The ALFWorld example trains "tokenizer_path=Qwen/Qwen2.5-3B-Instruct" through the scheduler and lists "`adv_estimator`: Advantage estimator (`gae`, `grpo`, `grpo_per_step`)".

• https://raw.githubusercontent.com/open-tinker/OpenTinker/main/README.md recorded 2026-10-19

  Quick Start table lists LLM single- and multi-turn math, an LLM single-LoRA math run, VLM single- and multi-turn geometry, and Gomoku, ALFWorld and Android World agents; an Environments section describes data-dependent and data-free, single- and multi-turn paradigms.