[whatmodel.sh](../../README.md) / GPT-NeoX-20B
# GPT-NeoX-20B
EleutherAI · released 2022-02 · 20.55B params · apache-2.0
**score 27.8** ±0.8 · #678 of 732 · from 32 benchmark results across 7 sources
Score within the [LMArena (text)](https://whatmodel.sh/?set=lmarena-text&score=set) comparison set only.
## What it's good at
● marks GPT-NeoX-20B among every model on that task or test, worst to best.
other tests (24)
test result vs the others in its run difficulty counts why
ARC Challenge 40.8% ▁·▁··▁▁▇▂●▆▃ beats 15/26 easy · avg 82% 34% sharp · several runs
WinoGrande 66.1% ▁····▁··▁▂●▄ beats 15/26 easy · avg 86% 11% at its level · several runs
HellaSwag 71.4% ▃···▁▇▅●▇▃▁▁ beats 14/26 very easy · avg 91% 7.4% at its level · several runs
ARC Easy 68.7% ▁▁·▁▁▄▅▄●▅▁▃ beats 15/26 easy · avg 85% 5.3% blunt · at its level · several runs
BoolQ 69.5% ▂▃▂▅▆▂●▆█▂▂▂ beats 12/26 easy · avg 88% 5.0% at its level
OpenBookQA 29.8% █·▁▁▃·▁▃●▇▃▄ beats 13/26 hard · avg 49% 4.9% too hard for it · several runs
SciQ 93.1% ▁·▁·▁·▃▅▃●▄▁ beats 14/26 very easy · avg 97% 4.5% blunt · several runs
PIQA 77.4% ▁·▁··▁▁▃█●█▂ beats 16/26 easy · avg 87% 3.6% blunt · at its level · several runs
TriviaQA (maker) 25.9% █···█·██●··█ beats 4/5 easy · avg 68% 3.2% at its level
LogiQA (harness) (maker) 23.0% ······●····· beats 4/5 very hard · avg 25% 3.0% too hard for it
LAMBADA (OpenAI version), last-wo… 72.0% ▁··▂▁▁▁▂▆█●▃ beats 20/26 very easy · avg 86% 2.9% blunt · at its level · several runs
CoNLL-2003 English NER, few-shot … 58.1% ▃···▅·▃●··▃█ beats 4/9 very easy · avg 85% 2.4% at its level
ANLI Round 1 (prompted) (maker) 34.0% █··●·······▂ beats 4/5 hard · avg 46% 2.0% sharp · too hard for it · noisy
Social IQA 42.0% ▁···▁▃●▁▁▁·▁ beats 11/26 hard · avg 56% 1.4% blunt
12-dataset text classification, 4… 56.6% ▂▂··▂·●···▂▂ beats 3/8 easy · avg 79% 1.3% blunt · at its level
ANLI Round 3 (prompted) (maker) 35.4% ▃······▃●··▃ beats 3/5 hard · avg 43% 1.2% too hard for it
ANLI Round 2 (prompted) (maker) 34.3% ██·██●·····█ beats 4/5 very hard · avg 43% 1.1% too hard for it · noisy
Anthropic HH helpful preference c… 55.0% ●········█·█ beats 0/2 hard · avg 65% 1.0% small run
QA4MRE 2013 (harness) (maker) 36.3% ▄····●█▄···▄ beats 1/5 medium · avg 57% 1.0% blunt · at its level
Open PT-LLM Leaderboard average (… 28.3% ●▁▁▁▁▁▁▁▂█▅▁ beats 19/276 medium · avg 72% 0.9% noisy
PROST (harness, micro accuracy) (… 29.6% ▃······█···● beats 5/5 very hard · avg 35% 0.9% blunt · too hard for it
SuperGLUE WSC (prompted) (maker) 50.0% ●··········▄ beats 2/5 hard · avg 62% 0.6% blunt
UD English-GUM UPOS tagging, few-… 83.6% ▄·▄▄·······● beats 4/4 very easy · avg 92% 0.6% blunt
MenatQA (all factors), exact matc… 24.6% ▂····▂··▂·●▄ beats 6/8 medium · avg 36% 0.4% blunt
source results share of the score
Stability AI 10 55% ▇▇▇▇▇▇▇
GPT-NeoX-20B: An Open-Source Autoregres… 16 38% ▇▇▇▇▇
Prompting Language Models for Linguisti… 2 3% ▇
Generative Calibration for In-context L… 1 1% ▇
trlX: A Framework for Large Scale Reinf… 1 1% ▇
Open Portuguese LLM Leaderboard 1 1% ▇
MenatQA: A New Dataset for Testing the … 1 0% ▇
prior: pull toward the middle 0% ▇
weighted average = 27.8
Licences: Stability AI ([Apache-2.0](../../licenses/stablelm-original-eval-json-Apache-2.0.txt)) · GPT-NeoX-20B: An Open-Source Autoregressive Language Model ([CC BY 4.0](../../licenses/paper-2204.06745v1-CC-BY-4.0.txt)) · Prompting Language Models for Linguistic Structure ([CC BY 4.0](../../licenses/acl-2023-acl-long-367-CC-BY-4-0.txt)) · Generative Calibration for In-context Learning ([CC BY 4.0](../../licenses/acl-2023-findings-emnlp-152-CC-BY-4-0.txt)) · trlX: A Framework for Large Scale Reinforcement Learning from Human Feedback ([CC BY 4.0](../../licenses/acl-2023-emnlp-main-530-CC-BY-4-0.txt)) · Open Portuguese LLM Leaderboard ([Apache-2.0](../../licenses/hf-leaderboard-spaces-Apache-2.0.txt)) · MenatQA: A New Dataset for Testing the Temporal Comprehension and Reasoning Abilities of Large Language Models ([CC BY 4.0](../../licenses/acl-2023-findings-emnlp-100-CC-BY-4-0.txt))
## How to run it
### llama.cpp (recommended)
1. Install llama.cpp. Homebrew and winget installs name the commands llama-cli and llama-server instead. [Other ways to install](https://github.com/ggml-org/llama.cpp/blob/master/docs/install.md)
```sh
# macOS / Linux
curl -LsSf https://llama.app/install.sh | sh
# Windows
irm https://llama.app/install.ps1 | iex
```
2. Run GPT-NeoX-20B (Q4_K_M, needs ~13.1 GB, from [huggingface.co/mradermacher](https://huggingface.co/mradermacher/gpt-neox-20b-GGUF/tree/25e46ca846af35af8abb3eb364f7ef070c9b2783))
```sh
llama cli -hf mradermacher/gpt-neox-20b-GGUF:Q4_K_M
```
3. Or start a local server and open the address it prints (http://localhost:9931)
```sh
llama serve -hf mradermacher/gpt-neox-20b-GGUF:Q4_K_M
```
Other versions: Q8_0 (~21.9 GB) `llama cli -hf mradermacher/gpt-neox-20b-GGUF:Q8_0`
### LM Studio
1. Install LM Studio. Open the app once before using the lms command. [Download it (macOS / Windows / Linux)](https://lmstudio.ai/download)
2. Run GPT-NeoX-20B (Q4_K_M, needs ~13.1 GB, from [huggingface.co/mradermacher](https://huggingface.co/mradermacher/gpt-neox-20b-GGUF/tree/25e46ca846af35af8abb3eb364f7ef070c9b2783))
```sh
lms get https://huggingface.co/mradermacher/gpt-neox-20b-GGUF@Q4_K_M -y
```
3. Then run lms ls to see its name, and lms chat followed by that name
Other versions: Q8_0 (~21.9 GB) `lms get https://huggingface.co/mradermacher/gpt-neox-20b-GGUF@Q8_0 -y` · 4bit (~11.6 GB, Mac only) `lms get https://huggingface.co/nhe-ai/gpt-neox-20b-mlx-4Bit -y`
### Ollama
1. Install Ollama. [Download it (macOS / Windows)](https://ollama.com/download)
```sh
# Linux
curl -fsSL https://ollama.com/install.sh | sh
```
2. Run GPT-NeoX-20B (Q4_K_M, needs ~13.1 GB, from [huggingface.co/mradermacher](https://huggingface.co/mradermacher/gpt-neox-20b-GGUF/tree/25e46ca846af35af8abb3eb364f7ef070c9b2783))
```sh
ollama run hf.co/mradermacher/gpt-neox-20b-GGUF:Q4_K_M
```
Other versions: Q8_0 (~21.9 GB) `ollama run hf.co/mradermacher/gpt-neox-20b-GGUF:Q8_0`
### MLX (Mac)
1. Install MLX (Mac). Apple silicon Macs only, text only.
```sh
# with uv
uv tool install mlx-lm
# or
pip install mlx-lm
```
2. Run GPT-NeoX-20B (4bit, needs ~11.6 GB, Mac only, from [huggingface.co/nhe-ai](https://huggingface.co/nhe-ai/gpt-neox-20b-mlx-4Bit/tree/1bfe032513d6c728e4cb3b9dfca06dbbf97574b1))
```sh
mlx_lm.chat --model nhe-ai/gpt-neox-20b-mlx-4Bit
```
3. Or start an OpenAI-compatible server at http://localhost:8080/v1
```sh
mlx_lm.server --model nhe-ai/gpt-neox-20b-mlx-4Bit
```
### Other ways to run it
1. Run it as published (fp16, needs ~41.29 GB, from [huggingface.co/EleutherAI](https://huggingface.co/EleutherAI/gpt-neox-20b/tree/c292233c833e336628618a88a648727eb3dff0a7))
```sh
uv run --with torch --with 'transformers<5' --with accelerate python -c "from transformers import pipeline; print(pipeline('text-generation', model='EleutherAI/gpt-neox-20b', torch_dtype='auto', device_map='auto')('Once upon a time', max_new_tokens=40)[0]['generated_text'])"
```
Model repository and documentation: [huggingface.co/EleutherAI](https://huggingface.co/EleutherAI/gpt-neox-20b)
[interactive version](https://whatmodel.sh/model/gpt-neox-20b) · [JSON](https://whatmodel.sh/api?format=json&all=1)