Model
Model explorer

Zephyr 7B Beta

OPEN
Hugging Face · Zephyr family · released Oct 25, 2023

DPO-aligned Mistral-7B fine-tune that popularized distilled preference tuning; weights, training recipe (alignment-handbook), and data all open under MIT.

ReasoningCodingVisionFunction callingTool useAgentic
311.2
Elo · rank #344
Parameters
7.24B
Active params
7.24B (dense)
Context
8K tokens
Architecture
Decoder-only Transformer (Mistral-7B-v0.1 base), distilled SFT + Direct Preference Optimization on synthetic feedback data
License
MIT
Languages
—
API price (in/out)
No hosted API
Modalities
text
Benchmark results
Bar shows position within the tracked field; marker = field best
AlpacaEvalHuman preference90.6%#3
best: Tülu 2+DPO 70B · 95.1%
ARC-ChallengeReasoning62.0%#81
best: Llama 3.1 405B · 96.9%
DROPReasoning9.66#66
best: Hunyuan-T1 · 93.1
GSM8KMath12.7%#174
best: Llama 3.1 405B · 96.8%
HellaSwagReasoning84.4%#49
best: Claude 3 Opus · 95.4%
MMLUKnowledge61.1%#201
best: OpenAI o3 · 92.9%
MT-BenchHuman preference7.34#55
best: Hunyuan-Large (A52B) · 9.4
TruthfulQAKnowledge57.5%#32
best: Phi-3.5-MoE (16x3.8B, 6.6B active) · 77.5%
WinoGrandeReasoning77.7%#45
best: PaLM 2 · 90.9%
Run it locally
VRAM @ Q4
5 GB
VRAM @ FP16
15 GB
Fits on (Q4)
RTX 3060 12GBRTX 4070 Ti 16GBRTX 3090 24GBRTX 4090 24GBRTX 5090 32GBM4 Pro 48GBM3 Max 128GBM3 Ultra 512GBA100 80GBH100 80GBH200 141GBB200 192GB
Throughput data unavailable.
Quantizations
GGUF Q4 · AWQ · GPTQ
Zephyr family
Elo progression across releases
Zephyr 7B BetaOct 2023311.2
Resources
HFHuggingFaceH4/zephyr-7b-betaGHhuggingface/alignment-handbookDOCSarxiv.org/abs/2310.16944
API price weights · each benchmark row carries its own source badge (see methodology)