Zephyr

weights

open(MIT

data

open(UltraChat_200k + UltraFeedback_binarized, both MIT, ungated)

code

open(alignment-handbook DPO recipe, Apache-2.0)

license

MIT(OSI)

orpo-checkpoint

open(zephyr-orpo-141b-A35b-v0.1, a separately released ORPO checkpoint, ships under Apache-2.0

Genuinely fully open: MIT weights, both DPO training datasets public and MIT-licensed, and end-to-end training code (the alignment handbook) under Apache-2.0 that reproduces zephyr-7b-beta. One caveat that does not change the class: the upstream UltraChat and UltraFeedback sets were partly generated with OpenAI model outputs, though the released Hugging Face artifacts are themselves open and MIT.

• https://github.com/huggingface/alignment-handbook recorded 2026-08-13

  Apache-2.0; contains all training code to replicate zephyr-7b-beta

• https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k recorded 2026-09-29

  Dataset page keywords include "mit", "100K - 1M" and the card says the original data "consists of 1.4M dialogues generated by ChatGPT and spanning a wide range of topics".

• https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized recorded 2026-09-29

  Dataset page keywords: "text-generation", "English", "mit", "100K - 1M"; description says "GPT-4 is then used to assign a score to each completion".

• https://huggingface.co/HuggingFaceH4/zephyr-7b-beta recorded 2026-09-29

  Model card repo state lists "license":"mit" and "gated":false; the card says the model was "trained on on a mix of publicly available, synthetic datasets using" DPO, with the datasets HuggingFaceH4/ultrachat_200k and HuggingFaceH4/ultrafeedback_binarized named.

• https://huggingface.co/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1 recorded 2026-09-29

  The 141B-A39B variant page lists License: "apache-2.0" and describes "synthetic, high-quality, multi-turn preferences that have been scored via LLMs".

• https://raw.githubusercontent.com/huggingface/alignment-handbook/main/LICENSE recorded 2026-09-29

  The repository LICENSE file is the full text of the Apache License, opening with "Apache License" and "Version 2.0, January 2004", and includes the notice "Licensed under the Apache License, Version 2.0".

• https://raw.githubusercontent.com/huggingface/alignment-handbook/main/README.md recorded 2026-09-29

  The alignment-handbook README announces "We release all the training code to replicate Zephyr-7b-β" and points to the recipe instructions to "Replicate Zephyr-7b-β by following the" recipe; its citation block lists license = {Apache-2.0}.

Adoption is measured as Hugging Face downloads of the two declared Zephyr models, almost all of it for the 7B flagship; the 141B ORPO variant sees negligible use. The UltraChat and UltraFeedback alignment datasets the product also declares are not added to the model figure.

• https://huggingface.co/api/datasets/HuggingFaceH4/ultrachat_200k recorded 2026-08-13

  73,327 downloads in the trailing 30 days for HuggingFaceH4/ultrachat_200k

• https://huggingface.co/api/datasets/HuggingFaceH4/ultrafeedback_binarized recorded 2026-08-13

  16,375 downloads in the trailing 30 days for HuggingFaceH4/ultrafeedback_binarized

• https://huggingface.co/api/models/HuggingFaceH4/zephyr-7b-beta recorded 2026-10-07

  87,178 downloads in the trailing 30 days for HuggingFaceH4/zephyr-7b-beta

• https://huggingface.co/api/models/HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1 recorded 2026-10-07

  224 downloads in the trailing 30 days for HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1

Zephyr-7B-beta was strong for a 7B model at its late-2023 release, with its card reporting strong results on several MT-Bench categories against larger open models such as Llama2-Chat-70B. It is weak by current standards (GSM8K 12.74, DROP 9.66 on the leaderboard) and sits well below current frontier chat models.

• https://huggingface.co/HuggingFaceH4/zephyr-7b-beta recorded 2026-09-29

  Card text: "In particular, on several categories of MT-Bench, Zephyr-7B-β has strong performance compared to larger open models like Llama2-Chat-70B"; the table lists "GSM8K (5-shot)" and "DROP (3-shot)" among leaderboard results.