AI Potluck
Model components / Fine-tuned / chat models

Tulu 3

Ai2

Ai2's open post-training model family at 8B, 70B and 405B, applying a fully reproducible recipe on top of Meta's Llama 3.1 base: open supervised and preference data, the open-instruct training code, and reinforcement learning with verifiable rewards. The evaluation framework is released alongside it.

The recipe, data and code are Ai2's own and reproducible, while the released weights inherit the base model's terms, which the openness axis follows. Verified 2026-08-13 via the model cards, the Tulu 3 blog and the open-instruct repository.

Openness

3 high confidence
3.0
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 - not OSI-approved, with a 700M monthly-active-user commercial cap - and that license governs what someone downloading this checkpoint is bound by, whatever Ai2 layered on top of it. The post-training layer itself is genuinely open: data, code, RLVR recipe and evaluation framework, all Apache-2.0, and it is credited where it lives - the tulu-3-sft-mixture corpus scores as fully open in its own right among the training and synthetic datasets, and the same recipe produces Olmo 3 Instruct at fully open on an open base. Scoring the method on the checkpoint would score two things at once. A license that bounds commercial use rather than forbidding it caps at 3, in the open-weights class.

  • https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B recorded 2026-08-13

    `license:llama3.1` in the repo tags, so the checkpoint ships under the Llama 3.1 Community License Agreement rather than Ai2's own terms; `"gated":false`, so the weights download without a barrier; base meta-llama/Llama-3.1-8B

  • https://allenai.org/blog/tulu-3 recorded 2026-08-13

    fully-transparent post-training release -- open data + data mixes, training code/recipe, eval framework, and weights

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

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

Adoption

2 high confidence
2.0

Research-scale reach. The three declared model SKUs sum to 10,342 downloads in the trailing 30 days (Llama-3.1-Tulu-3-8B 9,421; -70B 680; -405B 241), which bands at level 2 (10K-100K) on the model scale. The tulu-3-sft-mixture dataset the product also declares reads 36,267, but a product bands on the sum within one kind of artifact and the Hugging Face models win here, so the corpus is scored as its own product among the training and synthetic datasets rather than added in.

Capability

3 high confidence
3.0

A 2025 post-training recipe on a Llama 3.1 base, at the top of its own generation and below the 2026 frontier. The 405B card's results table gives the Tulu 3 405B column MMLU (5 shot, CoT) 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 / 68.8 - a comparison set now two model generations old, which is what keeps this at 3 rather than higher.

Verified 2026-08-13