Zephyr
Hugging Face's 2023 DPO-tuned Mistral 7B, now a reference model
Overview
Zephyr-7B-β is a 2023 chat model from Hugging Face's H4 team: Mistral-7B fine-tuned on UltraChat and aligned with distilled DPO on UltraFeedback, which beat Llama2-Chat-70B on MT-Bench at release. The MIT-licensed weights remain free to download, but the model has not been updated since 2024 and has no named successor.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Zephyr
- MIT-licensed 7B chat weights, free to download and run locally or via Ollama
- A reproducible distilled-DPO recipe in the alignment-handbook repository
- The Zephyr technical report (arXiv 2310.16944) documenting the method
- A 2023 baseline for comparing newer small-model alignment
- A worked example for teaching supervised fine-tuning and preference optimisation
Zephyr-7B-β earned its place in October 2023: a 7B model whose MT-Bench score beat Llama2-Chat-70B, with the whole distilled-DPO recipe published in the alignment-handbook. The MIT-licensed weights are still ungated and downloadable, and it remains a clean example for learning how preference tuning works.
As a model to deploy in 2026, skip it. It has not been updated since 2024, comes with no hosted service or support from Hugging Face, and its card warns it was not safety-aligned and can produce problematic text. Treat it as a teaching and research reference.
What we found
Zephyr-7B-β is a chat model released by Hugging Face's H4 team in October 2023, and this page treats it as what it now is: a historically important checkpoint rather than a current product. The team took Mistral AI's Mistral-7B-v0.1, fine-tuned it on the UltraChat 200k dataset, then aligned it with Direct Preference Optimization on UltraFeedback, a set of 64,000 prompts with completions ranked by GPT-4. The accompanying paper, Zephyr: Direct Distillation of LM Alignment, called the method distilled DPO.
The result mattered because of the comparison. The model card reports an MT-Bench score of 7.34 against 6.86 for Llama2-Chat-70B, a model ten times its size, and calls Zephyr the highest-ranked 7B chat model on MT-Bench and AlpacaEval at release. On AlpacaEval the 70B model still scored higher, 92.66% to 90.60%. The full training recipe was published in H4's alignment-handbook repository, and that openness, more than the checkpoint itself, is the lasting contribution.
The weights are MIT-licensed, ungated and still downloadable, and Ollama still lists zephyr in 7B and 141B sizes. The repository has not changed since October 2024. On Hugging Face's inference routing, one provider, Featherless AI, is mapped to serve it today. The same team published two later variants in 2024: zephyr-7b-gemma, under Google's Gemma terms, and Zephyr 141B-A35B, an ORPO tune of Mixtral-8x22B under Apache 2.0. Hugging Face names no successor. Its current small open models are the separate SmolLM family; SmolLM3-3B arrived in July 2025.
Use Zephyr to study alignment, reproduce the paper or teach fine-tuning, not to deploy. The card states that H4 removed the in-built alignment of its datasets to raise benchmark scores and that the model was not aligned for safety in an RLHF phase, so it can produce problematic text when prompted. It has had no updates since 2024 and comes with no support or hosted service from Hugging Face.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Zephyr-7B-Beta scored higher than Llama 2 70B on MT-Bench helpfulness metrics despite being 10x smaller — demonstrating that alignment quality matters more than model size for perceived helpfulness.
- The model was trained in just 1 GPU-week using DPO, compared to months of RLHF training required by comparable traditional models.
- Zephyr's success in late 2023 triggered a wave of DPO-aligned open-source models and validated DPO as the new standard alignment technique for open-source LLMs.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Zephyr's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Zephyr unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Zephyr directly →
Spotted something out of date? Suggest an update →
Related reading
Alternatives to Zephyr
More in LLMs & Chat
