GPT-NeoX · 9B to 32B
GPT NeoX 20B
20.74 billion parameters across 44 layers. At four-bit precision the weights alone come to 11.66 GB, before any conversation is loaded. This one is cache-hungry: a 32k window adds 33.0 GB on top, so context length matters more here than the model size suggests.
Memory needed, by quantisation
At an 8k context window. Quality is our rough ranking of how much the compression costs you: anything at or above Q5 is hard to tell apart from the original in normal use.
| Quantisation | File size | KV cache | Total VRAM | Quality | What it costs you |
|---|---|---|---|---|---|
| Q8_0 | 20.52 GB | 8.25 GB | 29.75 GB | 99% | Lossless in practice. Use it when the memory is there. |
| Q6_K | 15.84 GB | 8.25 GB | 25.06 GB | 98% | Very close to Q8 for two thirds of the size. |
| Q5_K_M | 13.74 GB | 8.25 GB | 22.96 GB | 96% | The quality-first choice when Q6 will not fit. |
| Q4_K_M | 11.66 GB | 8.25 GB | 20.89 GB | 93% | The default. Best size-to-quality ratio for local use. |
| IQ4_XS | 10.26 GB | 8.25 GB | 19.49 GB | 91% | Importance-matrix 4-bit. Q4_K_S quality, smaller file. |
| Q3_K_M | 9.44 GB | 8.25 GB | 18.67 GB | 86% | Degradation starts to show. A way to fit one size up. |
| IQ3_XXS | 7.39 GB | 8.25 GB | 16.61 GB | 79% | Aggressive. Only worth it on very large models. |
What a longer conversation costs
Same model at Q4_K_M, only the context window changes. The cache grows in a straight line with every token in the window.
| Context | KV cache | Total VRAM | Fits in 8 GB | Fits in 12 GB | Fits in 24 GB |
|---|
The "fits" columns allow for the roughly 0.8 GB Windows keeps for the desktop.
Which hardware runs GPT NeoX 20B
50 of 118 consumer devices run it at a quantisation worth using, at an 8k context window. Another 24 can load it only by compressing the weights far enough to damage the model, marked with a warning below.
| Device | Memory | Verdict | Quantisation | Used | Speed |
|---|---|---|---|---|---|
| GeForce RTX 5090 NVIDIA | 32 GB | Runs great | Q8_0 | 29.75 GB | 51.1 t/s |
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Runs great | Q8_0 | 29.75 GB | 51.1 t/s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Runs great | Q8_0 | 29.75 GB | 27.4 t/s |
| Apple M3 Ultra 96GB Apple | 96 GB | Runs well | Q8_0 | 29.45 GB | 22.2 t/s |
| Apple M3 Ultra 256GB Apple | 256 GB | Runs well | Q8_0 | 29.45 GB | 22.2 t/s |
| Apple M3 Ultra 512GB Apple | 512 GB | Runs well | Q8_0 | 29.45 GB | 22.2 t/s |
| RTX A6000 NVIDIA | 48 GB | Runs well | Q8_0 | 29.75 GB | 21.9 t/s |
| Apple M1 Ultra 64GB Apple | 64 GB | Runs well | Q8_0 | 29.45 GB | 21.7 t/s |
| Apple M1 Ultra 128GB Apple | 128 GB | Runs well | Q8_0 | 29.45 GB | 21.7 t/s |
| Apple M2 Ultra 64GB Apple | 64 GB | Runs well | Q8_0 | 29.45 GB | 21.7 t/s |
| Apple M2 Ultra 128GB Apple | 128 GB | Runs well | Q8_0 | 29.45 GB | 21.7 t/s |
| Apple M2 Ultra 192GB Apple | 192 GB | Runs well | Q8_0 | 29.45 GB | 21.7 t/s |
| Radeon PRO W7900 AMD | 48 GB | Runs well | Q8_0 | 29.75 GB | 21.6 t/s |
| GeForce RTX 4090 NVIDIA | 24 GB | Just fits | Q5_K_M | 22.96 GB | 37.6 t/s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Just fits | Q5_K_M | 22.96 GB | 37.6 t/s |
| GeForce RTX 3090 NVIDIA | 24 GB | Just fits | Q5_K_M | 22.96 GB | 34.9 t/s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Just fits | Q5_K_M | 22.96 GB | 33.4 t/s |
| Radeon RX 7900 XTX AMD | 24 GB | Just fits | Q5_K_M | 22.96 GB | 31.4 t/s |
| RTX A5000 NVIDIA | 24 GB | Just fits | Q5_K_M | 22.96 GB | 28.6 t/s |
| Apple M4 Max 36GB Apple | 36 GB | Just fits | Q6_K | 24.76 GB | 17.7 t/s |
| Apple M4 Max 48GB Apple | 48 GB | Runs well | Q8_0 | 29.45 GB | 14.8 t/s |
| Apple M4 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 29.45 GB | 14.8 t/s |
| Apple M4 Max 128GB Apple | 128 GB | Runs well | Q8_0 | 29.45 GB | 14.8 t/s |
| Apple M3 Max 36GB Apple | 36 GB | Just fits | Q6_K | 24.76 GB | 13 t/s |
| Apple M1 Max 32GB Apple | 32 GB | Runs well | Q4_K_M | 20.59 GB | 15.7 t/s |
| Apple M2 Max 32GB Apple | 32 GB | Runs well | Q4_K_M | 20.59 GB | 15.7 t/s |
| Apple M1 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Apple M2 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Apple M2 Max 96GB Apple | 96 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Apple M3 Max 48GB Apple | 48 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Apple M3 Max 64GB Apple | 64 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Apple M3 Max 96GB Apple | 96 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Apple M3 Max 128GB Apple | 128 GB | Runs well | Q8_0 | 29.45 GB | 10.8 t/s |
| Radeon RX 7900 XT AMD | 20 GB | Just fits | Q3_K_M | 18.67 GB | 32.6 t/s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Runs well | Q5_K_M | 22.96 GB | 10.2 t/s |
| Apple M4 Pro 48GB Apple | 48 GB | Runs well | Q4_K_M | 20.59 GB | 10.7 t/s |
| Apple M4 Pro 64GB Apple | 64 GB | Runs well | Q4_K_M | 20.59 GB | 10.7 t/s |
| Ryzen AI Max+ 395 32GB AMD | 32 GB | Fits, but slow | Q4_K_M | 20.89 GB | 8.4 t/s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Fits, but slow | Q4_K_M | 20.89 GB | 8.4 t/s |
| Ryzen AI Max+ 395 64GB AMD | 64 GB | Fits, but slow | Q8_0 | 29.75 GB | 5.8 t/s |
| Ryzen AI Max+ 395 96GB AMD | 96 GB | Fits, but slow | Q8_0 | 29.75 GB | 5.8 t/s |
| Ryzen AI Max+ 395 128GB AMD | 128 GB | Fits, but slow | Q8_0 | 29.75 GB | 5.8 t/s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Fits, but slow | Q8_0 | 29.75 GB | 5.8 t/s |
| Apple M1 Pro 32GB Apple | 32 GB | Fits, but slow | Q4_K_M | 20.59 GB | 7.8 t/s |
| Apple M2 Pro 32GB Apple | 32 GB | Fits, but slow | Q4_K_M | 20.59 GB | 7.8 t/s |
| Apple M3 Pro 36GB Apple | 36 GB | Just fits | Q6_K | 24.76 GB | 4.9 t/s |
| Apple M4 Pro 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.31 GB | 13.6 t/s |
| Apple M5 32GB Apple | 32 GB | Fits, but slow | Q4_K_M | 20.59 GB | 6 t/s |
| Apple M4 32GB Apple | 32 GB | Fits, but slow | Q4_K_M | 20.59 GB | 4.7 t/s |
| Intel Core Ultra 9 288V 32GB Intel | 32 GB | Fits, but slow | Q4_K_M | 20.89 GB | 4.4 t/s |
| Ryzen AI 9 HX 370 32GB AMD | 32 GB | Fits, but slow | Q4_K_M | 20.89 GB | 4.2 t/s |
| Apple M5 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.31 GB | 7.6 t/s |
| Apple M4 24GB Apple | 24 GB | Just fits | IQ3_XXS ! | 16.31 GB | 6 t/s |
| GeForce RTX 5080 NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 59.5 t/s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 55.6 t/s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 47.6 t/s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 45.6 t/s |
| GeForce RTX 4080 NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 44.5 t/s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 41.7 t/s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Runs great | IQ2_XXS ! | 14.2 GB | 35.7 t/s |
What to buy to run GPT NeoX 20B
The cheapest hardware that runs it at a quantisation worth using and a speed you would not resent, at an 8k context window.
Cheapest that works
GeForce RTX 3090
$700 used
Q5_K_M · uses 22.96 GB of its 24 GB · about 34.9 tokens/s
Best value
GeForce RTX 3090 Ti
$800 used
Q5_K_M · about 37.6 tokens/s · 1008 GB/s
If it just has to run
Radeon RX 7900 XT
$650
Q3_K_M · about 32.6 tokens/s. Cheaper because it runs the model at a lower precision.
| Device | Price | Memory | Runs it at | Speed | |
|---|---|---|---|---|---|
| GeForce RTX 3090 NVIDIA | $700 used | 24 GB | Q5_K_M | 34.9 t/s | Check price |
| GeForce RTX 3090 Ti NVIDIA | $800 used | 24 GB | Q5_K_M | 37.6 t/s | Check price |
| Apple M1 Max 32GB Apple · whole machine | $1,400 | 32 GB | Q4_K_M | 15.7 t/s | Check price |
| Apple M1 Max 64GB Apple · whole machine | $1,900 | 64 GB | Q4_K_M | 15.7 t/s | Check price |
| GeForce RTX 5090 NVIDIA | $2,200 | 32 GB | Q8_0 | 51.1 t/s | Check price |
| Apple M1 Ultra 128GB Apple · whole machine | $3,400 | 128 GB | Q8_0 | 21.7 t/s | Check price |
| Apple M3 Ultra 256GB Apple · whole machine | $5,599 | 256 GB | Q8_0 | 22.2 t/s | Check price |
| Apple M3 Ultra 512GB Apple · whole machine | $9,499 | 512 GB | Q8_0 | 22.2 t/s | Check price |
Indicative prices reviewed 2026-09-01; used prices are marketplace typical. Options that cost more than a cheaper one with no more memory and no more speed are hidden. Price links are Amazon affiliate links. Change the standard, the context window or the budget in the buying tool.
Where it sits in the catalogue
Source: EleutherAI/gpt-neox-20b on Hugging Face. Downloaded 787k times in the last month. Published 2022-04-07. Architecture figures are read from the repository's own configuration file, so they move when the model does.