Atlas

Models

← All models

StableLM 2 Zephyr 1.6B

Stability AI · 2024-01-19 · 1.6B parameters

StableLM 2 Zephyr 1.6B is Stability AI's compact 1.6B instruction-tuned language model from the StableLM 2 series, trained using Direct Preference Optimization (DPO) on a combination of public and synthetic datasets following the Zephyr training recipe. Released in January 2024, it targets strong instruction-following performance at a very small model size suitable for on-device or resource-constrained deployment. Despite its size, the model achieves competitive performance on chat benchmarks relative to other sub-2B models.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)33.5
Capability74.7
EQ-Bench v215.0
EQ-Bench v2 + MAGI-Hard Combined21.3
GPQA (Open LLM Leaderboard v2)24.3
IFEval (Open LLM Leaderboard v2)32.8
MAGI-Hard27.5
MATH Level 5 (Open LLM Leaderboard v2)3.3
MMLU-Pro (Open LLM Leaderboard v2)17.1
MuSR (Open LLM Leaderboard v2)35.1
Loading Atlas data…