jadidbourbaki/time-wizard

Time Wizard

Time Wizard is a 450 million parameter vision language model that reads the time from a photograph of an analog clock. Time Wizard answers with the hour and the minute as JSON. On 200 held out photographs Time Wizard matches GPT-5.6 Sol at maximum reasoning effort and beats Claude Fable 5.1 and Claude Opus 5 by a wide margin.

Property	Value
Base model	LFM2.5-VL-450M by Liquid AI
Parameters	450 million
Training data	time-wizard-bench train split
Training	full fine-tune, 142 seconds on one H100
Within 1 min on the test split	59.0%
Input	one photograph cropped to the clock
Output	{"hours": H, "minutes": M}
Licence	LFM Open License v1.0

We built Time Wizard by fine-tuning Liquid AI's LFM2.5-VL-450M on 2796 labelled clock photographs. The training run took 142 seconds on one H100. The photographs, labels, and splits are the public dataset time-wizard-bench. The training and scoring code is in the time-wizard repository on GitHub.

Results

The test split of time-wizard-bench holds 200 clock crops. No model saw them before scoring. Every model received the same 448 pixel crop and the same question. GPT-5.6 Sol, Claude Fable 5.1, and Claude Opus 5 ran through their APIs at the max reasoning effort setting with a budget of 32000 tokens per clock. Time Wizard decodes greedily with no reasoning.

Model	Effort	Within 1 min	Exact	Hour	Mean error	Unreadable
Time Wizard	none	59.0%	31.0%	74.5%	44 min	0
GPT-5.6 Sol	max	61.5%	31.0%	68.5%	65 min	0
Claude Fable 5.1	max	21.5%	11.5%	42.5%	107 min	2
Claude Opus 5	max	15.5%	6.0%	31.0%	141 min	0

Accuracy in the figure is the share of the 200 test clocks each model read correctly. The dark bars count a reading as correct within one minute of the label. The light bars count a reading as correct when the hour matches. The error bars extend one standard error above and below each share. The standard error is the amount a share would typically move if we drew a fresh 200 clocks from the same pool. Near 60 percent it is about 3.5 points. Two bars whose error bars overlap are within noise of each other.

Within 1 min is the share of clocks read to within one minute on the 12 hour dial. The distance wraps at twelve. Exact is the share read to the minute. Hour is the share with the right hour. Mean error averages the wrapped distance in minutes over readable replies. Unreadable counts replies with no valid time in them. Fable 5.1 used its whole token budget on two clocks without answering. The time-wizard-bench card gives the prompt, the scoring code, and the full configuration of each frontier model.

With 200 photographs the standard error near 60 percent is about 3.5 points. Time Wizard and GPT-5.6 Sol are within sampling error of each other on the one minute metric. Time Wizard leads on the hour and on mean error. Time Wizard leads Fable 5.1 by 38 points and Opus 5 by 44 points on the one minute metric.

Three test clocks

Tower clock	Station clock	Blurred clock

Model	Effort	Tower clock	Station clock	Blurred clock
Label		3:28	1:20	1:00
Time Wizard	none	3:28	1:20	8:20
GPT-5.6 Sol	max	3:28	10:07	6:00
Claude Fable 5.1	max	4:25	1:23	7:55
Claude Opus 5	max	4:28	10:22	7:10

The tower clock and the station clock are sharp. Time Wizard reads both to the minute. GPT-5.6 Sol and Opus 5 swap the hour and minute hands on the station clock. The blurred clock covers a few dozen pixels in the source photograph. Every model guesses wrong on it.

Perception and reasoning

Reading an analog clock means judging the angle of two hands. One minute moves the minute hand six degrees. Reading to the minute therefore means resolving angles to six degrees. GPT-5.6 Sol, Fable 5.1, and Opus 5 approach the task by reasoning. They describe the hands in words, estimate the angles, check which hand is longer, and verify the arithmetic. Each clock costs them thousands of tokens. The answers still land far from the label. Opus 5 misses by 141 minutes on average. A random guess would miss by 180.

Time Wizard reads the clock in one forward pass. The vision encoder of Time Wizard learned from 2796 labelled examples to carry the angle of each hand through to the language model. The language model turns the angles into a time. Time Wizard has 450 million parameters. Training it cost about fifteen cents of GPU time. On this task that is enough to match GPT-5.6 Sol.

The price gap at inference is large. We priced each API model's run by multiplying the input and output tokens it consumed by the list prices on the day, then dividing by the 200 clocks. GPT-5.6 Sol spent 686,146 output tokens on the 200 clocks. At its list price that is $14.04, or 7 cents a clock. Fable 5.1 spent 971,794 output tokens for $49.25, or 25 cents a clock. Opus 5 cost 4.7 cents a clock. We priced Time Wizard by its GPU time. Time Wizard read the 200 clocks in 26 seconds on one H100 that rents for $3.85 an hour on Nebius. The GPU time comes to 0.014 cents a clock.

Accuracy in the figure is the share of the 200 test clocks read within one minute of the label. The cost axis is the price per clock computed above, on a log scale because the four prices span three orders of magnitude.

Usage

The time-wizard repository provides a reader. Clone it, run just setup, then run:

just read clock.jpg

The reader downloads Time Wizard from the Hub and prints the time it reads, such as {"hours":3,"minutes":28}. Crop the photograph to the clock first. In Python, Reader loads Time Wizard once and reads many photographs:

from PIL import Image
from timewizard.reader import Reader

reader = Reader()
time = reader.read(Image.open("clock.jpg"))
print(time.hours, time.minutes)

read returns None when the reply holds no valid time. Reader pads the image to a square and resizes it to 448 pixels, the shape Time Wizard trained on.

To use the weights without the time-wizard repository, send one image and the fixed prompt through transformers:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
from PIL import Image, ImageOps

SYSTEM = "Reply with ONLY the requested JSON, no preface and no code block."
PROMPT = 'What time does this analog clock show? Reply as JSON: {"hours": H, "minutes": M} with H from 1 to 12.'

processor = AutoProcessor.from_pretrained("jadidbourbaki/time-wizard", max_image_tokens=256)
model = AutoModelForImageTextToText.from_pretrained("jadidbourbaki/time-wizard", dtype=torch.bfloat16)

image = ImageOps.pad(Image.open("clock.jpg").convert("RGB"), (448, 448), color=(0, 0, 0))
messages = [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM}]},
    {"role": "user", "content": [{"type": "image", "image": image}, {"type": "text", "text": PROMPT}]},
]
inputs = processor.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
)
out = model.generate(**inputs, max_new_tokens=32, do_sample=False)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[1] :], skip_special_tokens=True)[0])

The system prompt and the question must match the strings above. Time Wizard saw only those strings during training. max_image_tokens=256 is the image budget we trained with. A 448 pixel square occupies 196 image tokens under it. Greedy decoding with 32 new tokens is enough for the JSON reply.

Training

We fine-tuned every weight of LFM2.5-VL-450M on the 2796 photograph train split of time-wizard-bench. Each photograph is a crop around one clock with a human label for the hour and the minute. We turned each photograph into one conversation: the system prompt, the image with the question, and the label as the JSON reply. We then trained LFM2.5-VL-450M to predict the reply.

We used supervised fine-tuning with TRL's SFTTrainer. LFM2.5-VL-450M predicts each token of the reply. We penalise wrong predictions with cross entropy. We mask the system prompt, the image, and the question out of the loss. LFM2.5-VL-450M therefore spends its capacity on producing the time.

We trained the vision encoder along with the language model. The encoder has to change. Reading a clock requires the encoder to carry the angle of each hand through to the language model. A low rank adapter would have frozen the encoder. LFM2.5-VL-450M fits on one GPU with room to spare. Full fine-tuning therefore costs nothing extra.

We used AdamW with a peak learning rate of 5e-5. We warmed the rate up over the first three percent of steps and decayed it with a cosine schedule. Each step saw 32 photographs as two accumulated batches of 16. We trained for five epochs in bfloat16 with seed 0. Five epochs take about 440 steps. After every epoch we measured loss on the 200 photograph dev split of time-wizard-bench. We kept the epoch with the lowest dev loss. Epoch three reached the minimum of 0.249. The weights published here are the weights saved at the end of epoch three. Epochs four and five ran to completion. The epoch four and five checkpoints had a higher dev loss. We discarded them.

We chose the learning rate and the number of epochs from a sweep scored on the dev split. The rate mattered a great deal. At 2e-5 the fine-tune reached 28 percent within one minute after three epochs and 38 percent after five. At 5e-5 over three epochs the fine-tune collapsed to 1.5 percent. The same rate over five epochs reached 60.5 percent. The longer schedule stretches the warmup and slows the decay. The longer schedule is what makes the higher rate usable. Dev loss tracked the dev score across every run. The lowest dev loss therefore picked the best fine-tune without touching the test split. Two seeds at the same setting landed half a point apart. The differences above are real.

The final run took 142 seconds on one H100 on Nebius. At the hourly price of that machine the run cost about fifteen cents. The whole sweep of seven runs took under twenty minutes of GPU time.

Limitations

Time Wizard reads a clock that fills most of the image. Time Wizard has seen only crops with a margin of one fifth of the clock's width on each side. A whole scene with a small clock in it needs a detector or a manual crop first.

Many photographs cannot be read to the minute by anyone. Four in five COCO clocks cover under 224 pixels of the source photograph. At that size one minute of arc is under a pixel wide. On such clocks Time Wizard gets the hour right more often than the minute.

The dial gives no AM or PM. The hour is always between 1 and 12. A clock with no minute hand, a digital clock, or a stopped clock all produce a confident answer that may be wrong. We trained and evaluated on photographs of real clocks. We did not test drawings or rendered clocks.

The headline number rests on 200 photographs. A difference of a few points between two models on this test split is within noise.

Licence

Liquid AI releases LFM2.5-VL-450M under the LFM Open License v1.0. Time Wizard is a fine-tune of LFM2.5-VL-450M and carries the same licence. Read the licence before use.

Citation

@misc{timewizard2026,
  title  = {Time Wizard: A 450M Analog Clock Reading Model},
  author = {Tirmazi, Hayder},
  year   = {2026},
  url    = {https://huggingface.co/jadidbourbaki/time-wizard}
}