Nomm.ai
Salmon, grains, avocado, broccoli and tomatoes in a bowl

TWO AI MODELS TRAINED BY NOMMAI

Grape & Plum.
Our food AI.
Free.

Meet Grape XS and Plum S: two calorie and macro AI models trained by NommAI. Free to run on your device, with unlimited scans.

Open weights on Hugging Face. Grape for quick meal totals; Plum for a more detailed food breakdown.
No paid inference API. Your photos stay on your phone.

Meet Grape & Plum ↓
Download on the App Store
Get it on Google Play

On-device processing · Calorie tracking · Your food diary

Salmon Illustrative portion
Vegetables
Grains
✦Ready to explore● Local AI
✦ AI model estimateRECORDED RESULT
699
kcalEstimated calories
46g protein
48g carbs
39g fat
✓ Review foods & portions
0AI API fees
Local models stay free.
Recorded Grape 1 and Plum 1 estimates for this AI-generated image, tested on iPhone. Animation is illustrative; no live inference or photo upload. Estimates are not measured nutrition.

TRAINED BY NOMMAI

Two trained models.
Open weights. Free.

NommAI is available free on the App Store and Google Play. Grape and Plum provide on-device calorie and macro estimates, with different sizes and capabilities.

Grape 1 and Plum 1 weights are public under their published model licenses. They are free to run locally in NommAI. The research paper has been submitted to arXiv and is awaiting publication. We’ll link the methods and evaluation when the paper is public.

On-device inferencePublic model weightsResearch paper pending

Public model weights. Paper submitted to arXiv; public link pending. Available free on iOS and Android.

OUR COMMITMENT

Public model weights.
Free to use on your phone.

Use Grape and Plum on your phone as often as you want, for free. No scan credits, monthly quota or paid AI API subscription is required for these on-device models.

Grape is included in the app. Plum is an optional download of about 1.04 GB during setup or later in model settings. Public model licenses include usage restrictions. Larger models accessed over an API are planned as an optional service with separate terms. Your free on-device models will remain available.

A SIMPLE ROUTINE

From photo
to food diary.

You stay in control of the estimate.

Yogurt breakfast with berries, banana, granola and mintAI-generated food photograph
  1. 01
    ▣

    Take or pick a photo

    Use the camera or choose a meal from your photo library.

  2. 02
    ✦

    Run local AI

    Grape estimates meal totals on your phone. Plum adds a food-by-food breakdown on supported devices.

  3. 03
    ✓

    Review, then log

    Adjust foods and portions before saving. Track meals against your chosen calorie goal.

INSIDE NOMM.AI

Scan and review
your meals.

See your meal, check the estimate and save it to your diary. A clear view on iPhone. More room to review on iPhone Duo.

01

IPHONE DUO · EXPANDED DISPLAY

Your photo. Your estimate.
Side by side.

Native expanded iOS meal review: salmon photo beside the recorded 699 kcal Grape estimate and editable portions
Runs on your deviceCheck portions before saving
02

Review it.

Check calories, macros
and portions.

iPhone review screen showing the same salmon meal, 699 estimated kcal and macro cards
iPhone · Meal review
03

Keep it.

Your meal and photo,
saved in your diary.

iPhone diary with the same saved salmon meal and 699 logged kcal
iPhone · Food diary
Close-up of the salmon, avocado and vegetables used in the app demonstrationYOUR MEAL PHOTO

PHOTO → ESTIMATE → DIARY

The photo stays with the meal.

Review the food you photographed, adjust what you know and keep the picture with your saved entry.

Explore the app
Colorful breakfast bowl with berries, banana and granolaBREAKFAST, LUNCH OR DINNER

Native iOS simulator captures in illustrated device frames. The salmon photo is AI-generated; the displayed values are recorded estimates, not measured nutrition. Expanded-layout previews do not establish physical-device performance.

Local tracking does not require an account. First setup records your notice acknowledgements under a random installation ID. Diary sync and photo sharing stay off unless you choose them.

TRAINED BY NOMMAI · OPEN WEIGHTS

Grape & Plum.

Choose quick meal totals or a food-by-food breakdown. Both models are trained by NommAI. Both are free, run on your phone and have public weights on Hugging Face.

TRAINED BY NOMMAIFREE · OPEN WEIGHTS
Grape AI model emblem

Grape

Whole-meal estimates

XSEXTRA SMALL

A compact AI model for everyday logging. Get estimated calories and macros for your whole meal, then review and save.

Quick totalsOur lighter on-device model
  • Calories, protein, carbs & fat
  • Included in the app
  • Photos stay on your phone
Grape XS · Weights & license ↗Trained by NommAI
TRAINED BY NOMMAIFREE · OPEN WEIGHTS
Plum AI model emblem

Plum

Food-by-food estimates

SSMALL

More detail from the same photo. Combine meal-total estimation with individual food identification, then check the portions before saving.

A longer, more detailed scanWhole-meal totals + food breakdown
  • Individual foods & estimated portions
  • Download once, then scan offline
  • Supported phones with 6 GB+ RAM
Plum S · Weights & license ↗Trained by NommAI
IN DEVELOPMENT

Larger models are planned.

Future additions to the NommAI model family.

NommAI Plus

Coming soon
M

NommAI Pro

Coming soon
L

XS → S → M → L describes size, not measured accuracy. Speed varies by phone, photo and model version. Available free on iOS and Android. Plum can take several minutes on some Android phones. Plus and Pro are previews; availability, pricing and hosting are not announced.

✳

ESTIMATES, EXPLAINED

A photo cannot
measure everything.

Portion size, cooking oils and hidden ingredients affect calorie estimates. Review the result, make corrections, or add another photo when needed. Calorie goals are estimates too, and you can set your own.

PlateBenchView PlateBench

GRAPE & PLUM EVALUATIONS

Small models.
Full comparison.

See how our food AI compares.

Compare Grape and Plum with other models on meals photographed at home, in restaurants and in controlled tests.

View PlateBench on GitHub
6/8

Lowest average error

Plum research results · 6 of 8 tests
10,021

Photo evaluations

Eight tests across six datasets
188 MB

Grape model size

Plum combines Grape + a 0.8B model

Published research results. Performance on your phone and individual meals can differ.

MODEL COMPARISON

The complete comparison.

Average calorie error across eight food-photo tests.
Lower numbers are better. Highlighted cells show the lowest error.

Average calorie error (kcal). Lower is better. Highlighted cells show the lowest reported error, not statistical significance.
Food-photo test
Plum 11.2 GB
Grape 1188 MB
Qwen3.8-27B27B
J50.8B
CalorieK94M
Boba0.8B
Nutrition5k overhead507 photosLowest point estimate: 45.350.192.749.8104.9108.5
Nutrition5k angled675 photosLowest point estimate: 43.948.7102.747.457.283.3
NutritionVerse889 photos490.3525.5Lowest point estimate: 427.7462.3657.2680.7
SNAPMe1,474 photosLowest point estimate: 131.2136.6135.8133.7158.0197.5
JFB1,000 photosLowest point estimate: 131.6143.9132.2135.2178.5241.4
DiningBench1,650 photosLowest point estimate: 219.0232.0226.0227.2312.7472.2
DiningBench customer1,458 photosLowest point estimate: 231.0239.6236.5243.4320.3458.1
OmniFood8K2,368 photos173.0167.9Lowest point estimate: 143.0189.7228.1259.7

Scores show the average difference from the dataset’s reference calories (kcal MAE). Green cells mark the lowest reported error for each test. Grape and Plum scores are research averages; see test details below.

About the tests and results

Grape and Plum results are two-seed research recipe averages from PlateBench, computed on a PC. They are not direct measurements of the shipped phone models. Qwen3.8-27B is a zero-shot server baseline; J5 is our research calorie-decision component. CalorieK and Boba are research baselines. Earlier development candidates are available in the source tables.

Dataset references vary: some meals are weighed, some use self-reported food records or corrected labels, and restaurant labels can be model-assisted. Related photos are not all independent meals. Some public tracks informed model choices, so this table is not an untouched holdout. Lowest average errors do not establish statistical superiority or guarantee an accurate estimate for every meal.

PlateBench is maintained by us. The repository includes separate held-out results, confidence intervals, predictions and scoring code. The research paper’s arXiv link is pending.

Read the full results and methods ↗

Explore our models.Public weights. Free on-device scans.

Grape weights ↗Plum weights ↗

Source snapshot: PlateBench b217f5e · Checked 2026-10-27 · Download chart data (JSON) · Artwork attribution

NO SUBSCRIPTION FOR LOCAL AI

Other apps charge.
Your phone runs it free.

Grape and Plum run on your device. Scan as often as you want, review your estimates and keep your food diary without a subscription or per-scan fee.

Our model weights are public under their published licenses. The research paper has been submitted to arXiv and is awaiting publication, so you’ll be able to read the methods and evaluation when it is public.

Why local AI stays free
UNLIMITED ON-DEVICE SCANS$0

No trial countdown. No scan credits. Your free local models stay available.

Larger models over an API are planned as an optional service. They won’t be required to use Grape or Plum locally.

THE DETAILS

Before you start.

Is NommAI free?

Yes. Grape 1 and Plum 1 run on your phone for free, with unlimited local scans and no paid AI API subscription. Their weights are public under licenses with usage restrictions. Larger API-based models and optional cloud services are planned separately; no subscription starts when you install or accept the notices.

How accurate are photo estimates?

Calories and portions are estimates. Oils, sauces and hidden ingredients can be missed. Review the foods and amounts, correct the numbers, or enter a meal manually before saving.

Do you upload my meal photos?

On-device scans and saved photos stay on your phone by default. Optional cloud backup, hosted reviews and support sharing are being tested separately and require your choice. We do not automatically collect photos for training.

Can I use NommAI offline?

Installed models analyze photos on your device. Internet is required to confirm your first acceptance of the notices, download Plum and acknowledge later required policy updates. Reading your existing diary remains available offline.

Which AI model does NommAI use?

Grape 1 estimates whole-meal nutrition; Plum 1 adds food-by-food estimates on supported phones. Both are made by NommAI. Public download files and licenses are available; the research paper has been submitted to arXiv and its public link is pending.

When can I download the app?

NommAI is available free on the App Store. It is also available free on Google Play. Grape and Plum weights are public. See download availability. The paper link will follow when its abstract is public.

NOW ON IOS & ANDROID

Get NommAI.
Start with Grape.

Our own trained AI models. Free, unlimited calorie and macro estimates on your phone.

Grape XS is included with the app. Add optional Plum S for food-by-food details on supported devices. No subscription or account required for local tracking.

Download on the App Store
Get it on Google Play

Choose your store and download the app. Start with the included Grape model for the quickest scans.

See model sizes and device guidance ↗

Meal tracking.
On your terms.

Free photo calorie tracking with on-device AI and public model weights.

Download on the App Store
Get it on Google Play
Download NommAI ↗