Tulu 3

weights

open(downloadable)

license

Llama-3.1-Community(non-OSI

post-training-data

open(Apache-2.0/ODC-BY)

code

open(open-instruct Apache-2.0)

recipe

open(SFT->DPO->RLVR)

evals

open(OLMES)

The released weights inherit Meta's Llama 3.1 Community License, non-OSI and capped at 700 million monthly active users for commercial use, governing the checkpoint regardless of what Ai2 adds. The post-training layer above it is fully open under Apache-2.0: the data, the open-instruct code, the SFT-DPO-RLVR recipe and the OLMES eval framework, which keeps this at open weights rather than open source.

• https://allenai.org/blog/tulu-3 recorded 2026-09-29

  Ai2 describes Tulu 3 as a release of "fully open-source data, code, and recipes."

• https://github.com/allenai/open-instruct recorded 2026-08-13

  Apache-2.0; Tulu 3 SFT/DPO/RLVR post-training framework + decontamination tools

• https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B recorded 2026-09-29

  Model card tags include "license:llama3.1" and the card names the "Llama 3.1 Community License Agreement".

• https://huggingface.co/datasets/allenai/tulu-3-sft-mixture recorded 2026-09-29

  The SFT mixture card states "ODC-BY-1.0 license; different licenses apply to subsets of the data."

• https://raw.githubusercontent.com/allenai/open-instruct/main/LICENSE recorded 2026-09-29

  The open-instruct LICENSE file is the "Apache License" text, "Version 2.0, January 2004".

• https://raw.githubusercontent.com/allenai/open-instruct/main/README.md recorded 2026-09-29

  The open-instruct README says the repo releases "Code for DPO, preference finetuning and reinforcement learning with verifiable rewards (RLVR)." and points to OLMES, "which we used for TÜLU 3".

Adoption sums Hugging Face downloads across Tulu's three declared model checkpoints, 8B, 70B and 405B, most of the volume on the 8B. The tulu-3-sft-mixture dataset the product also declares is scored separately as its own entry among the training and synthetic datasets, rather than added to the model count.

• https://huggingface.co/api/models/allenai/Llama-3.1-Tulu-3-405B recorded 2026-10-07

  341 downloads in the trailing 30 days for allenai/Llama-3.1-Tulu-3-405B

• https://huggingface.co/api/models/allenai/Llama-3.1-Tulu-3-70B recorded 2026-10-07

  586 downloads in the trailing 30 days for allenai/Llama-3.1-Tulu-3-70B

• https://huggingface.co/api/models/allenai/Llama-3.1-Tulu-3-8B recorded 2026-10-07

  6,422 downloads in the trailing 30 days for allenai/Llama-3.1-Tulu-3-8B

Tulu 3 is a 2025 post-training recipe on a Llama 3.1 base. Its 405B results table gives MMLU 87.0, GSM8K 95.5, IFEval 86.0, HumanEval 95.9 and MATH 67.3, against a GPT-4o (11-24) column at 87.9, 91.7, 84.8, 97.0 and 68.8, and it sits below the current frontier.

• https://huggingface.co/allenai/Llama-3.1-Tulu-3-405B recorded 2026-09-29

  The 405B card table has rows "MMLU (5 shot, CoT)" and "GSM8K (8 shot, CoT)"; Tülu 3 is "designed for state-of-the-art performance on a diversity of tasks in addition to chat".