EleutherAIGPTNeoXForCausalLM

GPT Neox 20B — Hardware Requirements & GPU Compatibility

Chat

GPT Neox 20B is a 20.7B-parameter open language model from EleutherAI. It supports a context window of up to 2,048 tokens. At Q4_K_M it needs about 13.69 GB of VRAM — see which GPUs and Macs can run it below.

20.7B params · dense2K contextApache 2.0Apr 2022258.0K downloads587 likes5.4K quant downloads

Your hardware

Verdicts below are for the reference RTX 4090. Set yours to personalise the page.

Detection reads your browser's GPU name with no data sent anywhere.

Runs well on a 24 GB GPU

Runs from 8 GB (IQ2_M · 2K)Best on 24 GB (Q4_K_M · ~48 tok/s · 2K)Macs from 16 GB (IQ4_XS · ~11 tok/s)

Apache 2.0

Reference GPU · RTX 4090 24 GB

Estimates, ±20% · how we estimate

Run GPT Neox 20B

One recommended way. Every other method, OS and quant is in the step-by-step guide.

Other ways

each opens step-by-step

Run in cloud

Fits on RTX 3090 (24 GB) (10 GB headroom) · Q4_K_M

Generation speed: ~45 tok/s; generation speed
Cost per 1M output tokens: $1.35; per 1M output tokens

Deployon Vast.aiBest optionfrom $0.22/hr

Compare GPUs →

Run GPT Neox 20B

13.7 GB VRAM at 2K context · fits the RTX 4090 (10.3 GB free)

Works on macOS, Windows and Linux.

Install Ollama (Linux)
$curl -fsSL https://ollama.com/install.sh | sh
Install Ollama (macOS / Windows)
Download and run the app; it installs the ollama command.
Download Ollama ↗
Run Q4_K_M
$ollama run hf.co/mradermacher/gpt-neox-20b-i1-GGUF:Q4_K_M
Raise the context window
Type it inside the chat. For the API server instead, start it with OLLAMA_CONTEXT_LENGTH=2048 ollama serve.
$/set parameter num_ctx 2048
A larger context needs more memory than the weights alone.
Use the API
Ollama serves its API on http://localhost:11434.

Works on macOS, Windows and Linux.

Install LM Studio
Desktop app for macOS (Apple silicon), Windows and Linux.
lmstudio.ai ↗
Download the model
In the app press Ctrl/Cmd + Shift + M and paste https://huggingface.co/mradermacher/gpt-neox-20b-i1-GGUF, then pick Q4_K_M. Or use the CLI (run LM Studio once first):
$lms get https://huggingface.co/mradermacher/gpt-neox-20b-i1-GGUF@Q4_K_M
Open on Hugging Face ↗
Load and chat
Open the model loader, select the model and set Context Length to 2048. Leave GPU offload at maximum if it fits.
Start the local server (optional)
$lms server start

Works on macOS, Windows and Linux.

Install llama.cpp (macOS)
MacPorts, Nix and conda-forge also work.
$brew install llama.cpp
Install options ↗
Install llama.cpp (Windows)
$winget install llama.cpp
Install llama.cpp (Linux)
Download the prebuilt llama-*-bin-ubuntu-* archive for your GPU backend (CPU, Vulkan) from the releases page, unpack it and run the binaries from that folder. Homebrew (brew install llama.cpp) and conda-forge also work on Linux.
llama.cpp releases ↗
Start the server
Downloads the model on first run. The chat UI and an OpenAI-compatible API are at http://localhost:8080.
$llama-server -hf mradermacher/gpt-neox-20b-i1-GGUF:Q4_K_M -c 2048 -ngl 99 --port 8080

Works on macOS, Windows and Linux.

Option A: Docker Model Runner
Docker Desktop: Settings > AI > Enable Docker Model Runner. Docker Engine on Linux: install the docker-model-plugin package.
$docker model run hf.co/mradermacher/gpt-neox-20b-i1-GGUF:Q4_K_M
Model Runner docs ↗
The quant tag after the repo is how Hugging Face cards write it; if your Docker version rejects it, drop the tag and pick the quant by repo.
Option B: llama.cpp server image (NVIDIA)
Needs the NVIDIA Container Toolkit. Models are cached in the llama-cache volume. Open http://localhost:8080.
$docker run --gpus all -p 8080:8080 -v llama-cache:/root/.cache ghcr.io/ggml-org/llama.cpp:server-cuda -hf mradermacher/gpt-neox-20b-i1-GGUF:Q4_K_M -c 2048 -ngl 99 --host 0.0.0.0 --port 8080
Image tags (cuda13, vulkan, rocm) ↗
Cache path inside the image is not in the docs; remove the -v flag if the volume complains.

Linux only.

Install vLLM
Linux only; on Windows run it inside WSL2. NVIDIA GPU shown; AMD ROCm, Intel and CPU builds have their own install pages.
$uv venv --python 3.12 --seed
$source .venv/bin/activate
$uv pip install vllm --torch-backend=auto
vLLM installation ↗
Serve the model
OpenAI-compatible API at http://localhost:8000/v1.
$vllm serve EleutherAI/gpt-neox-20b --max-model-len 2048
Runs the original weights, not the Q4_K_M GGUF; needs ~46 GB of GPU memory.

macOS only.

MLX supports only macOS, so follow these steps for that setup.

Install mlx-lm
$pip install mlx-lm
Chat in the terminal
$mlx_lm.chat --model nhe-ai/gpt-neox-20b-mlx-4Bit
Downloads the model on first run. The MLX repo has its own quantisation; it may not match the quant selected above.
Or start a local server
Listens on http://localhost:8080.
$mlx_lm.server --model nhe-ai/gpt-neox-20b-mlx-4Bit

Works on macOS, Windows and Linux.

Install Jan
jan.ai ↗
Find the model
Open the Hub tab, paste this model ID in Search, and choose Q4_K_M if offered.
$mradermacher/gpt-neox-20b-i1-GGUF
Use the local API (optional)
Settings > Local API Server > Start Server. It listens on http://127.0.0.1:1337.

Rent GPUs hourly to run Q4_K_M right there. Pricing comes from partner marketplaces and varies.

Available GPU rentals for Q4_K_M, prioritized by value
GPU	Speed	$/1M tokens	Rent
RTX 3090 (24 GB)CheapestCheapest10 GB free	~45 tok/s	$1.35	Deploy on Vast.aifrom $0.22/hr (opens in new tab)	24 GB · 10 GB free · ~45 tok/s · $1.35 / 1M tokens
RTX A5000 (24 GB)10 GB free	~37 tok/s	$1.77	Deploy on Vast.aifrom $0.23/hr (opens in new tab)	24 GB · 10 GB free · ~37 tok/s · $1.77 / 1M tokens
TITAN RTX (24 GB)10 GB free	~32 tok/s	$2.11	Deploy on Vast.aifrom $0.24/hr (opens in new tab)	24 GB · 10 GB free · ~32 tok/s · $2.11 / 1M tokens
RTX PRO 4000 (24 GB)10 GB free	~32 tok/s	$2.92	Deploy on Vast.aifrom $0.33/hr (opens in new tab)	24 GB · 10 GB free · ~32 tok/s · $2.92 / 1M tokens
RTX PRO 4500 (32 GB)18 GB free	~43 tok/s	$2.22	Deploy on RunPodfrom $0.34/hr (opens in new tab)	32 GB · 18 GB free · ~43 tok/s · $2.22 / 1M tokens
RTX 4090 (24 GB)10 GB free	~48 tok/s	$1.97	Deploy on RunPodfrom $0.34/hr (opens in new tab)	24 GB · 10 GB free · ~48 tok/s · $1.97 / 1M tokens
A40 (48 GB)34 GB free	~33 tok/s	$2.95	Deploy on RunPodfrom $0.35/hr (opens in new tab)	48 GB · 34 GB free · ~33 tok/s · $2.95 / 1M tokens
L4 (24 GB)10 GB free	~14 tok/s	$8.61	Deploy on RunPodfrom $0.44/hr (opens in new tab)	24 GB · 10 GB free · ~14 tok/s · $8.61 / 1M tokens

Compare GPUs →

How Much VRAM Does GPT Neox 20B Need?

Select a quantization level; the listed hardware adheres to it.

VRAM	+32K	+128K	File	Quality	On RTX 4090 (reference)
6.3 GB	—	—	5.5 GB	2-bit · i-quant extreme compression — Importance-weighted 2-bit, extreme compression — significant quality loss	Fits: 2K context · ~105 tok/s
6.8 GB	—	—	6.1 GB	2-bit · i-quant extra small — Importance-weighted 2-bit, extra small	Fits: 2K context · ~96 tok/s
7.1 GB	—	—	6.5 GB	2-bit · i-quant small — Importance-weighted 2-bit, small	Fits: 2K context · ~92 tok/s
7.7 GB	—	—	7 GB	2-bit · i-quant medium — Importance-weighted 2-bit, medium	Fits: 2K context · ~85 tok/s
8.8 GB	—	—	8.1 GB	3-bit · i-quant — Importance-weighted 3-bit	Fits: 2K context · ~74 tok/s
9.4 GB	—	—	8.7 GB	3-bit · i-quant extra small — Importance-weighted 3-bit, extra small	Fits: 2K context · ~70 tok/s
9.7 GB	—	—	9 GB	3-bit · i-quant small — Importance-weighted 3-bit, small	Fits: 2K context · ~68 tok/s
9.7 GB	—	—	7.8 GB	2-bit · K-quant — 2-bit quantization with K-quant improvements	Fits: 2K context · ~68 tok/s
10.0 GB	—	—	9 GB	3-bit small — 3-bit small quantization	Fits: 2K context · ~66 tok/s
10.3 GB	—	—	10 GB	3-bit · i-quant medium — Importance-weighted 3-bit, medium	Fits: 2K context · ~64 tok/s
11.1 GB	—	—	10.8 GB	3-bit medium — 3-bit medium quantization	Fits: 2K context · ~59 tok/s
11.4 GB	—	—	11.7 GB	4-bit · legacy — 4-bit legacy quantization	Fits: 2K context · ~57 tok/s
11.7 GB	—	—	11.8 GB	3-bit large — 3-bit large quantization	Fits: 2K context · ~56 tok/s
12.3 GB	—	—	11 GB	4-bit · i-quant compact — Importance-weighted 4-bit, compact	Fits: 2K context · ~53 tok/s
12.8 GB	—	—	11.8 GB	4-bit small — 4-bit small quantization	Fits: 2K context · ~51 tok/s
13.7 GB	—	—	13.1 GB	4-bit medium — 4-bit medium quantization — most popular sweet spot	Fits: 2K context · ~48 tok/s
15.7 GB	—	—	14.2 GB	5-bit small — 5-bit small quantization	Fits: 2K context · ~42 tok/s
16.3 GB	—	—	15.3 GB	5-bit medium — 5-bit medium quantization — good quality/size tradeoff	Fits: 2K context · ~40 tok/s
18.8 GB	—	—	16.9 GB	6-bit, very good — 6-bit quantization, very good quality	Fits: 2K context · ~35 tok/s
22.8 GB	—	—	21.9 GB	8-bit, near-lossless — 8-bit quantization, near-lossless	Fits: 2K context · ~29 tok/s
45.6 GB	—	—	~41.5 GB	16-bit · full — Full half-precision — baseline for inference	Does not fit: needs 45.6 GB

est.calculated, no published file yet

Selected Q4_K_M

4-bit medium quantization

13.7GBat 2K context

Weights: 12.9 GB
KV cache: 0.5 GB
Overhead: 0.3 GB

Fits RTX 4090 (reference) · 10.3 GB free · ~48 tok/s

≥50 tok/s needs ≥1.05 TB/s · RTX 5090 ~85

Which Hardware Can Run GPT Neox 20B?

Q4_K_M · 13.7 GB

GPT Neox 20B (Q4_K_M) requires 13.7 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 18+ GB is recommended. 26 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 5080. 124 devices with unified memory can run GPT Neox 20B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M5 Pro (24 GB).

By VRAM tier · typical GPUs and Macs at Q4_K_M

Typical hardwareQuantVRAMtok/sContext that fits

8 GB · RTX 4060IQ2_M · 7.7 GB · ~23 tok/s · 2K7.7~232K8 GB tier →

12 GB · RTX 3060 12GBQ3_K_L · 11.7 GB · ~20 tok/s · 2K11.7~202K12 GB tier →

16 GB · RTX 5060 Ti 16GBQ4_K_M · 13.7 GB · ~21 tok/s · 2K13.7~212K16 GB tier →

24 GB · RTX 4090referenceQ4_K_M · 13.7 GB · ~48 tok/s · 2K13.7~482K24 GB tier →

Mac 32 GB · Mac Mini M4Q4_K_M · 13.7 GB · ~6 tok/s · 2K13.7~62KMacs →

Mac 64 GB · Mac Studio M4 MaxQ4_K_M · 13.7 GB · ~28 tok/s · 2K13.7~282KMacs →

How good is GPT Neox 20B?

llmrun Score blends public benchmark boards into one number; 100 = the average open model.

Not enough data yet for an llmrun Score.

A Score needs results on enough independent boards. How the Composite Score is built

Benchmarks

HellaSwag
Epoch AI · Apr 2022Lab-reported
70.5%
#34/47 open
Lab-reported
ARC Challenge
Epoch AI · Apr 2022Lab-reported
41.1%
#41/56 open
Lab-reported
WinoGrande
Epoch AI · Apr 2022Lab-reported
66.1%
#31/51 open
Lab-reported
BoolQ
Epoch AI · Apr 2022Lab-reported
64.9%
#32/38 open
Lab-reported
OpenBookQA
Epoch AI · Apr 2022Lab-reported
38.8%
#20/26 open
Lab-reported
PIQA
Epoch AI · Apr 2022Lab-reported
76.7%
#29/40 open
Lab-reported

All 6 GPT Neox 20B benchmarks

Alternatives to GPT Neox 20B

Similar models

4tok/s on RTX 4090

Model	Score	VRAM	tok/s	Why it is similar
InternVL3 5 GPT OSS 20B A4B Preview HFOpenGVLab	—	13.6 GB	~180	same size
Internlm 20BInternLM	—	45.3 GB	—needs > 24 GB	same size
ERNIE 4.5 21B A3B ThinkingBaidu	—	14.3 GB	~213	same size
Mistral Small Instruct 2409Mistral AI	—	17.4 GB	~46	same size

Specifications

Publisher: EleutherAI
Parameters: 20.7B
Architecture: GPTNeoXForCausalLM
Context Length: 2,048 tokens
Vocabulary Size: 50,432
Release Date: 2022-04-07
License: Apache 2.0

Resources

Hugging Face model card (opens in a new tab)

Frequently Asked Questions

How much VRAM does GPT Neox 20B need?

GPT Neox 20B requires 13.7 GB of VRAM at Q4_K_M, or 45.6 GB at FP16.

VRAM = Weights + KV Cache + Overhead

Weights = 20.7B × 4.8 bits ÷ 8 = 12.4 GB

KV Cache + Overhead ≈ 1.3 GB (at 2K context + ~0.3 GB framework)

Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.

VRAM usage by quantization

Q4_K_M

13.7 GB

Learn more about VRAM estimation →

Can NVIDIA GeForce RTX 4090 run GPT Neox 20B?

Yes, at Q8_0 (22.8 GB) or lower. Higher quantizations like FP16 (45.6 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.

What's the best quantization for GPT Neox 20B?

For GPT Neox 20B, Q4_K_M (13.7 GB) offers the best balance of quality and VRAM usage. Q5_K_S (15.7 GB) provides better quality if you have the VRAM. The smallest option is IQ2_XXS at 6.3 GB.

VRAM requirement by quantization

IQ2_XXS

6.3 GB~53%

IQ3_XS

9.4 GB~73%

Q3_K_M

11.1 GB~83%

Q4_K_M ★

13.7 GB~89%

Q5_K_S

15.7 GB~92%

FP16

45.6 GB~100%

★ Recommended — best balance of quality and VRAM usage.

Learn more about quantization →

Can I run GPT Neox 20B on a Mac?

Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run GPT Neox 20B, the smallest at IQ4_XS. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 13.7 GB of usable unified memory (RAM minus macOS overhead), from MacBook Pro 14" M3 Pro (18 GB).

Can I run GPT Neox 20B locally?

Yes — GPT Neox 20B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.

How fast is GPT Neox 20B?

At Q4_K_M, GPT Neox 20B can reach ~351 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~48 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.

tok/s = (bandwidth GB/s ÷ model GB) × efficiency

Example: NVIDIA B200 → 8000 ÷ 13.7 × 0.65 = ~380 tok/s

Estimated speed at Q4_K_M (13.7 GB)

NVIDIA B200

~380 tok/s

NVIDIA GeForce RTX 4090

~48 tok/s

NVIDIA B300

~380 tok/s

AMD Instinct MI350X

~351 tok/s

Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.

Learn more about tok/s estimation →

What's the download size of GPT Neox 20B?

At Q4_K_M, the download is about 12.44 GB. The full-precision FP16 version is 41.48 GB. The smallest option (IQ2_XXS) is 5.70 GB.

Which GPUs can run GPT Neox 20B?

26 consumer GPUs can run GPT Neox 20B at Q4_K_M. Top options include AMD Radeon RX 7900 XTX, NVIDIA GeForce RTX 3090, NVIDIA GeForce RTX 3090 Ti, AMD Radeon RX 6800. 7 GPUs have plenty of headroom for comfortable inference.

Which devices can run GPT Neox 20B?

124 devices with unified memory can run GPT Neox 20B at Q4_K_M (13.7 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.