Will it fit?

Pick a model and how many bits each of its numbers gets. See the download, the memory it needs, and where it can run.

Phone
Fits
8 GB memory, 60 GB/s
Needs 4% of its memory, leaving room for the system and your apps.
Writes at most 223 tokens a second: faster than you read.
Laptop
Fits
16 GB memory, 100 GB/s
Needs 2% of its memory, leaving room for the system and your apps.
Writes at most 372 tokens a second: faster than you read.
Data center
1 GPU
80 GB and 3,350 GB/s per GPU, 8 per server
One GPU holds it.
Writes at most over 1,000 tokens a second: faster than you read.

Model

Parameters135 million

Each number stored in

32-bit16-bit8-bit4-bit

Download269 MB

At 50 Mbit/s43 seconds

Memory to run323 MB

This site really downloads272 MB

Challenges

0 of 3 complete

Get Llama 2 7B to fit comfortably on the phone, using at most half of its memory. (not yet)
With 16-bit numbers, find the biggest model the laptop could still run faster than you read (at least 2 billion parameters). (not yet)
Fit Llama 3.1 405B into a single 8-GPU server. (not yet)

How it is worked out. Download = parameters times bytes per number (2 bytes at 16-bit); scales and small files add a few percent. Memory = download times 1.2, for working space (EleutherAI's rule of thumb; long conversations need more). Speed at best = memory speed divided by download size, because writing each token reads nearly every number once; real software is slower. File sizes count 1 GB as a billion bytes; memory is sold in binary gigabytes, so an "8 GB" phone holds 8.6 billion bytes.

Devices. 8 GB of memory, as in Google's Pixel 9a. Memory speed assumed to be 60 GB/s: phone makers rarely publish it. 16 GB of memory, the smallest MacBook Air sold in October 2026. Memory speed 100 GB/s: Apple's M2, the chip this site's own speeds were measured on. Each GPU (the SXM version of NVIDIA's H100) has 80 GB of memory at 3,350 GB/s, and a server such as NVIDIA's DGX H100 holds 8. Bigger models are split across GPUs that read their parts at the same time.

Try this

Start with SmolLM2-135M at 16-bit and predict its download before you look: 135 million parameters times 2 bytes. Then compare with what this site really downloads.
Choose Llama 2 7B and step from 32-bit down to 4-bit. Watch the phone and the laptop change.
At 16-bit, drag Parameters up slowly and watch the laptop’s best-case speed fall.
Choose Llama 3.1 405B and count the GPUs at each precision.

The same formula against the 15 models this site runs in your browser: the formula column is parameters times bits each, divided by 8 to get bytes. The real downloads are a little larger: they include small files such as the tokenizer, 8-bit versions keep a few small parts at 32-bit, and 4-bit files also store a scale for each small block of weights. Janus-Pro-1B stores its parts at different precisions, so its formula adds them up part by part, and the calculator above leaves it out.

Model	Parameters (millions)	Bits each	Formula	Real download
DeiT-Tiny Names what is in a photo	5.9	32	23.6 MB	24 MB
YOLOS-Tiny Finds and boxes objects in a photo	6.5	32	26.0 MB	27 MB
all-MiniLM-L6-v2 Turns sentences into numbers	22.7	8	22.7 MB	24 MB
Depth Anything V2 Small Estimates how far away things are	24.8	16	49.6 MB	50 MB
ViTPose+ small Finds a body's keypoints	33.1	16	66.3 MB	67 MB
SAM 2.1 tiny Outlines the object you click	39.0	16	77.9 MB	78 MB
BiRefNet lite Cuts the subject out of a picture	44.3	16	88.7 MB	99 MB
Supertonic 2 Reads text aloud	66.0	32	264 MB	263 MB
Whisper base.en Turns speech into text	72.6	8	72.6 MB	80 MB
Audio Spectrogram Transformer Names the sounds in a clip	86.6	8	86.6 MB	91 MB
SmolLM2-135M Writes text	134.5	16	269 MB	272 MB
SmolLM2-135M-Instruct Chats	134.5	16	269 MB	272 MB
CLIP ViT-B/32 Matches photos and words	151.3	16	303 MB	306 MB
OWL-ViT B/32 Finds anything you name in a photo	153.2	16	306 MB	311 MB
Janus-Pro-1B Makes pictures from text	1,530.7	4 to 32	1.02 GB	1105 MB
Janus-Pro-1B, reading pictures Answers questions about a photo, sharing two files with the row above	2,172.4	4 to 16	1.51 GB	1628 MB

The calculator in a few linesPython with the standard library

# Download, memory, and best-case writing speed on a laptop, from a parameter count.
GB = 1e9                           # file sizes: a billion bytes
LAPTOP_MEMORY = 16 * 2**30         # a "16 GB" laptop holds 16 x 2^30 bytes
LAPTOP_SPEED = 100 * GB            # bytes its memory moves per second (Apple M2)

def on_laptop(params, bits):
    download = params * bits / 8   # bytes: 2 per parameter at 16-bit, half at 4-bit
    memory = 1.2 * download        # plus about 20% working space
    share = memory / LAPTOP_MEMORY
    if share > 1:
        return download, memory, "too big"
    fit = "fits" if share <= 0.5 else "tight"
    speed = LAPTOP_SPEED / download  # each token reads every number once
    return download, memory, f"{fit}, at most {speed:.0f} tokens a second"

models = [("SmolLM2-135M", 134_515_008), ("Llama 2 7B", 6_738_415_616),
          ("Llama 3 70B", 70_553_706_496)]
for name, params in models:
    for bits in (16, 4):
        download, memory, verdict = on_laptop(params, bits)
        print(f"{name:12} {bits:2}-bit: download {download / GB:6.2f} GB, "
              f"memory {memory / GB:6.2f} GB: {verdict}")

SmolLM2-135M 16-bit: download   0.27 GB, memory   0.32 GB: fits, at most 372 tokens a second
SmolLM2-135M  4-bit: download   0.07 GB, memory   0.08 GB: fits, at most 1487 tokens a second
Llama 2 7B   16-bit: download  13.48 GB, memory  16.17 GB: tight, at most 7 tokens a second
Llama 2 7B    4-bit: download   3.37 GB, memory   4.04 GB: fits, at most 30 tokens a second
Llama 3 70B  16-bit: download 141.11 GB, memory 169.33 GB: too big
Llama 3 70B   4-bit: download  35.28 GB, memory  42.33 GB: too big

Phone

Laptop

Data center