AI Potluck
Model components / Fine-tuned / chat models

Zephyr

Hugging Face

Hugging Face H4's fully-open chat model line. The flagship zephyr-7b-beta is a DPO fine-tune of Mistral-7B, trained with the open alignment-handbook recipe on the public UltraChat and UltraFeedback datasets and relicensed MIT. A late-2023 model that remains a widely-downloaded, fully-reproducible 7B open baseline. Also includes the zephyr-orpo-141b MoE variant (Apache-2.0). Verified live on HF June 2026.

Scored on zephyr-7b-beta (MIT; base Mistral-7B Apache-2.0). Fully-open: weights + open DPO datasets (UltraChat/UltraFeedback, MIT) + open training code (alignment-handbook, Apache-2.0). Late-2023 capability, still ~165k HF downloads/month.

Openness

5 high confidence
5.0
weights
open(MIT
data
open(UltraChat_200k + UltraFeedback_binarized, both MIT, ungated)
code
open(alignment-handbook DPO recipe, Apache-2.0)
license
MIT(OSI)

Genuinely fully-open: MIT weights, the two DPO training datasets are public and MIT, and the end-to-end training code (alignment-handbook) is Apache-2.0 and replicates zephyr-7b-beta. Open_source tier 5. Caveat (does not change the class): upstream UltraChat/UltraFeedback were partly generated with OpenAI model outputs, but the released HF artifacts are open and MIT.

Adoption

4 high confidence
4.0

~164.9k HF downloads/month on zephyr-7b-beta with 1.85k likes -- a late-2023 model still pulling meaningful volume as a lightweight fully-open 7B baseline. The underlying datasets remain active (ultrachat_200k ~57.7k/mo). The 141B ORPO variant is effectively dormant (~52/mo). Level 4 on the leading SKU's sustained monthly volume.

Capability

2 high confidence
2.0

Notable late-2023 capability (a 7B beating much larger chat models on MT-Bench at release), but weak by 2026 standards (GSM8K 12.74, DROP 9.66 on the leaderboard). Low-tier on a scale where 5 = current frontier chat. Scored 2.

Unchanged since 2026-06-25 (last edited, not re-checked)