Zephyr 141B-A35B

Available · Language

Zephyr 141B-A35B is a Zephyr chat model from Hugging Face, fine-tuned from the Mixture of Experts model Mixtral 8x22B v0.1 and published as open weights under Apache 2.0. The card says Argilla, KAIST and Hugging Face trained it together to test ORPO at a larger scale; it gives 141B total and 39B active parameters, corrected from the 35B in the name. [1] [2]README, News: April 12, 2024 Primary source

Timeline of Zephyr 141B-A35B →

Claims and evidence

  • Released The model card was written on 2024-04-11 in a repository created on 2024-04-10; the handbook news gives April 12, 2024 as the release. Primary source[2]README, News: April 12, 2024 [1]Hub commit history
  • Status AvailableWeights are downloadable from the HuggingFaceH4 organisation on the Hub. Hugging Face publishes no deprecation page for its models. Primary source[1]Hub repository HuggingFaceH4/zephyr-orpo-141b-A35b-v0.1, Files and versions
  • Successor of Zephyr 7B GemmaThe card calls it the latest model in the Zephyr series; the handbook news lists Zephyr 7B Gemma as the preceding Zephyr release. Primary source[1]first paragraph [2]README, News
  • Derived from (fine tune) Mixtral 8x22B (Mistral AI) Primary source[1]
  • Change · Size A Mixture of Experts model with 141B total and 39B active parameters, against 7B for Zephyr 7B Gemma.Compared with Zephyr 7B Gemma Primary source[1]Model Description, Model type [3]Model description, Model type
  • Change · Architecture Built on the Mixture of Experts model Mixtral 8x22B v0.1 as base model instead of Gemma 7B.Compared with Zephyr 7B Gemma Primary source[1] [3]
  • Change · Other Aligned in one stage with ORPO, which needs no separate supervised fine-tuning step, on 7k synthetic multi-turn preference examples.Compared with Zephyr 7B Gemma Primary source[1]
  • Change · Licensing Released under the Apache 2.0 licence instead of the Gemma Terms of Use.Compared with Zephyr 7B Gemma Primary source[1]Model Description, License [3]Model description, License
  • Input text Primary source[1]first paragraph; Model Description
  • Output text Primary source[1]first paragraph; Model Description
  • Open weights YesApache 2.0 licence. Primary source[1]Model Description, License
  • Parameters 141B total parameters and 39B active parametersMixture of Experts. The model id and the first card version gave 35B active parameters; Hugging Face corrected this to 39B on 2024-04-18. Primary source[1]Model Description, Model type
  • Access Open-weights download Primary source[1]Intended uses and limitations

Lineage

Predecessors

Successors

No known successor.

Based on

  • Mixtral 8x22B · Mistral AI · 17 April 2024 · derived (fine tune)

Variants and derived

None recorded.

Siblings

None recorded.

All ancestors

All descendants

None.

Variants

No variants recorded in this record.

Related AI Radar coverage

AI Radar coverage starts in June 2026; no coverage linked yet.

All model releases from Hugging Face on AI Radar →